Exactly my experience, I know they vibe code features and that’s fine but it looks like they don’t do proper testing which is surprising to me because all you need bunch of cheap interns to some decent enough testing
No there is a wide gap between good and bad testers. Great testers are worth their weight in gold and delight in ruining programmer's days all day long.
IMO not a good place to skimp and a GREAT place to spend for talent.
I have little doubt where things are going, but the irony of the way they communicate versus the quality of their actual product is palpable.
Claude Code (the product, not the underlying model) has been one of the buggiest, least polished products I have ever used. And it's not exactly rocket science to begin with. Maybe they should try writing slightly less than 100% of their code with AI?
More generally, Anthropic's reliability track record for a company which claims to have solved coding is astonishingly poor. Just look at their status page - https://status.claude.com/ - multiple severe incidents, every day. And that's to say nothing of the constant stream of bugs for simple behavior in the desktop app, Claude Code, their various IDE integrations, the tools they offer in the API, and so on.
Their models are so good that they make dealing with the rest all worth it. But if I were a non-research engineer at Anthropic, I wouldn't strut around gloating. I'd hide my head in a paper bag.
I am constantly amazed how developers went hard for claude-code when there were and are so many better implementations of the same idea.
It's also a tool that has a ton of telemetry, doesn't take advantage of the OS sandbox, and has so many tiny little patch updates that my company has become overworked trying to manage this.
Its worst feature (to me at least), is the, "CLAUDE.md"s sprinkled all over, everywhere in our repository. It's impossible to know when or if one of them gets read, and what random stale effect, when it does decide to read it, has now been triggered. Yes, I know, I'm responsible for keeping them up to date and they should be part of any PR, but claude itself doesn't always even know it needs to update any of them, because it decided to ignore the parent CLAUDE.md file.
"Coding" is solved in the same way that "writing English language" is solved by LLMs. Given ideas, AI can generate acceptable output. It's not writing the next "Ulysses," though, and it's definitely not coming up with authentically creative ideas.
But the days of needing to learn esoteric syntax in order to write code are probably numbered.
On top of that is something they should have had from earlier times. My biggest pain point is to not to be able to continue from my phone. I just use a service to pipe telegram to any cc session in the dev machine. This is the number 1 reason why I got excited by openclaw in the first place but its overkill to have it just to control cc
That's a bummer. I was looking forward to testing this, but that seems pretty limiting.
My current solution uses Tailscale with Termius on iOS. It's a pretty robust solution so far, except for the actual difficulty of reading/working on a mobile screen. But for the most part, input controls work.
My one gripe with Termius is that I can't put text directly into stdin using the default iOS voice-to-text feature baked into the keyboard.
I’ve been doing this for a while [1], but ultimately settled on a building a thin transport layer for Telegram to accept and return media, and persistent channels, vastly improved messaging UX, etc. and ended up turning this into a ‘claw with a heartbeat and SOUL [2].
I really enjoyed reading both posts. Thanks for sharing!
I, like many others, have written my own "claw" implementation, but it's stagnated a bit. I use it through Slack, but the idea of journaling with it is compelling. Especially when combined with the recent "two sentence" journaling article[1] that floated through HN not too long ago.
Great posts! So far [2] is the only "claw" that has caught my interest, mostly because it isn't trying to do everything itself in some bespoke, NIH way.
I've been using email and Cloudeflare email router. You don't get the direct feedback of a terminal, but it's much easier to read what's happening in html formatted email.
It also feels kind of nice to just fire off an email and let it do it's thing.
I'm willing to bet most of their libraries are definitely vibe coded. I'm using the claude-agent-sdk and there are quite a few bugs and some weird design decisions. And looking through the actual python code it's definitely not what I would classify 'best practice'. Bunch of imports in functions, switching on strings instead of enums, etc.
I had to downgrade to an earlier release because an update introduced a regression where they weren't handling all of their own event types.
A few weeks ago the github integration was completely broken on the claude website for multiple days. It's very clear they vibe code everything and while it's laudable that they eat their own dogfood, it really projects a very amateurish image about their infrastructure and implementation quality.
> - You can't interrupt Claude (you press stop and he keeps going!)
This is normal behavior on desktop sometimes its in the middle of something? I also assume there's some latency
> - At best it stops but just keeps spinning
Latency issues then?
> - It can get stuck in plan mode
I've had this happen from the desktop, and using Claude Code from mobile before remote control, I assume this has nothing to do with remote control but a partial outage of sorts with Claude Code sometimes?
I don't work for Anthropic, just basing off my anecdotal experience.
I have just today discovered zmx [1] which is like tmux but I always hated the tmux terminal emulation and how it hijacks scrolling, especially on Termius on my phone. It does session persistence but I think without the terminal emulator side of things, so scrolling works normally.
Been testing it today with Claude Code and it seems to work quite well switching between my laptop and phone.
tmux supports tabs so you can have multiple Claude Code sessions running concurrently. You do need to learn a few tmux keyboard shortcuts to use it effectively (e.g. opening/closing/switching tab).
Yes. Doing the same. What is the advantage of this new feature? Tmux/Tailscale/Termius give you full control of your terminal.
Or mainly to save the end user the hassle to set it up correctly?
Oh lots of people will not be comfortable with tmux approach. The anthropic feature makes sense. But it's Max only and doesn't work well according to other comments.
Ease of setup is the biggest reason. I use this setup as well, but there are other UX niceties that would be a lot better with a dedicated mobile app: push notifications when Claude needs your input (I use a hook for this that connects to Pushover, but that's another service and extra setup), voice input, autocorrect that's right for this context, etc.
I used it to add a MIDI driver and support to my OS this afternoon. Worked okay, but I agree it is a bit clunky yet. I think it is pretty good for a preview release. Much better than nothing.
Opencode's 'web' command makes your local session run on the browser with same access rights as the cli. It's a pretty slick interface too. I sometimes use it instead of the cli even when I can access both.
You can test it right now if you want with the included free models.
It's changing super fast. I am using it on the desktop mostly and when I tried on my phone there were issues yes. But do try it out again in a few weeks.
(I am actually using zellij on the remote and using various CLIs more than I am using only opencode on the web. I was using wezterm mux until about a week ago but the current state of the terminal is not very good for this scenario. It seems like almost all the CLIs are choking because of nodejs ink library)
I was using this religiously but there’s a bug currently that makes the initialization fail and/or throws an error on the phone client.
Absolutely great piece of software otherwise, free, anonymous, encrypted and so on. Really hope the team can fix this soon - I would hate to switch back to tmux tunneling.
I feel like a lot of folks are saying this kills the Code on your Phone opportunity some start-ups are building for. I don't agree. I feel like coding agents are like streaming services, we will subscribe to multiple and switch between them. So for one there's value in a universal control plane. The other is that mobile as a coding interface should offer more than a remote control to the desktop. I think there's still some space to cook, especially if people are investing 8 hours a day talking to agents, the interface surely matters.
I don't know a single person who is satisfied with the status quo on streaming services where you have to subscribe to multiple ones. Everyone is complaining that the landscape is 1) more fragmented than cable was, 2) costs more, 3) has even more ads than cable
I think people forgot how bad it was. It was much more fragmented before but instead of services it was fragmented by time. Sure you have access to Seinfeld, but you can watch one or two Seinfelds a night at 8pm and 11pm.
I also remember base cable without any movies was around $60 or something and with some movie channels is >$100. And that's not inflation adjusted. You can easily get 3 or 4 of the top services for $100 today.
Finally claiming there are more ads on these services is a joke. There was ~20m for every 30m of programming, meaning 1/3 of the time you're watching commercials. And not just any commercials, the same commercials over and over. There was even a case of shows being sped up on cable to show more commercials.
I get it, everyone wants everything seamlessly and for next to nothing, but claiming that 90s cable was even comparable is absurd.
Not that it is particularly relevant to agentic coding but how can anyone truly argue streaming costs more? Average cable packages were exceeding 125-150 USD a month (in 2000 dollars). Under no circumstances would I be sympathetic to the argument that streaming costs more.
You can get all 7 of the major streaming subs for less without even shopping out deals. That is 100s of times the volume and quality of content that was delivered on cable for far less. It is so much content realistically that no one I have ever met has subscribed to all of them at once.
The argument really is empty. The fragmentized experience is annoying, but it isn't more expensive...And it DEFINITELY has fewer ads.
I agree. I spend a lot of time working from my phone so I had to make my own workflow that works for me. I've been following all these bans and drama with the subscription keys and custom harnesses etc. I think there's room for a "universal control plan" that lets you leverage the CLI providers (and whatever crappy interfaces / apis they give you).
Weird all these companies struggle so much to support remote services, ssh has been working for me pretty seamlessly for like the 20 years I've been using it and has allowed me to remote-control any computer I own with relatively reliable authentication (with some hiccups that tend to be patched pretty rapidly when found) throughout that entire period. I hear tell it worked even before I was using computers professionally, too
Claude Code is a good product, they should just keep on steadily improving it and improving the model. I am not sure why they are spraying in all directions like this..
Claude Code Team: Please fix the core experience instead of branching out into all these tertiary features. I know it is fun and profit to release new features but you need to go deeper into features not broader into there be dragons territory.
There is that (you can still use coding agents on the other 80% to polish, by the way) but to me this situation underscores the value of a good "editor"--someone who says this is good to ship vs. not this not now.
Maybe it’s related to what I tend to use the agents for, but I guess I don’t understand what is this for. Typically I try to structure the tasks in a way that require me to do or check something important when the agent gets back to me. If the agents query is trivial enough I can respond from my phone, it was likely not needed at all. If the agent finished - fine. It will have to wait until I get back in front of the computer anyway.
I've used similar things (omnara/happy) while taking walks. Sometimes I'll get an idea about the problem I'm working on and I can just dictate it into my phone and check in 15min later. I stopped being able to do that when claude added those nice interview panes to clarify things because it didn't work back then. But mostly it's really annoying when you think you've created the plan/prompt and that it's ready to go. But it gets stuck or decided to stop while you're away. I pretty often need to give Claude a "continue" kick. To be fair this happens far less after Opus 4.6.
Also, I felt the need to use it far more when I was on Pro vs a Max plan. On Pro when you hit the usage windows it's nice to be able to kick claude back into gear without scheduling your life around getting back to the terminal to type "continue".
- Plan mode -> answer questions/make corrections, continue planning
- Some of us don't do full yolo mode all the time, then tool approvals or code reviews are required, nice to do a quick review and decide if you need to go back to your computer or not
- Letting claude spin or handle a long-running task outside of normal work hours and being able to check in intermittently to see if something crashed
I don't dangerously accept permissions outside of a few scripts I have reviewed as safe. This means claude gets stuck often when testing it's work, but also means it doesn't uninstall production workloads from the kubernetes cluster.
Well it DOES have less storage than a Nomad (hence lame), but this way you don't need to pay for a public IP address, or for a VPS to run Wireguard on, or for a commercial VPN solution, and then install a terminal emulator on your phone and set up SSH keys.
People tried reinventing terminals, SSH, and tmux for phones. It's a pretty terrible experience using your thumbs. And it takes significant know-how to set up.
And in modern stacks, it almost necessitates a man in the middle - tailscale is common but it's still a central provider. So is it really the most inefficient way possible?
Fair point technically, but I think the value proposition isn't the persistent session, rathere it's the abstraction layer. Screen/tmux assumes you know what commands to run. This assumes you know what outcome you want. For someone like me who came to coding late and doesn't have 20 years of muscle memory with terminal tools, the inefficiency in transport is more than offset by the efficiency in intent. Different tools for different people.
Even better, train an entirely new LLM with your prompt added to its data set. It will be imbued with its own latent sense of purpose. All you need to do after that is type "let there be light!"
I'm probably 10 years out of date. Are ethereum smart contracts still a thing? I'm sure you could deploy one of those for every agent session to handle the notifications
Doesn't look like it has proper worktree management. UIs that abstract away worktrees are very powerful. I vibe coded my own (https://github.com/9cb14c1ec0/vibe-manager), which unfortunately doesn't have the remote component that hapi does.
My needs are very basic and it hasn’t failed me yet, I like that it doesn’t try to do much. I know it has voice capabilities through eleven labs but I haven’t used that feature.
I've been running something similar for a few months, which is a voice-first interface for Claude Code running on a local Flask server. Instead of texting from my phone, I just talk to it. It spawns agents in tmux sessions, manages context with handoff notes between sessions, and has a card display for visual output.
The remote control feature is cool but the real unlock for me was voice. Typing on a phone is a terrible interface for coding conversations. Speaking is surprisingly natural for things like "check the test output" or "what did that agent do while I was away."
The tmux crowd in this thread is right that SSH + tmux gets you 90% of the way there. But adding voice on top changes the interaction model. You stop treating it like a terminal and start treating it like a collaborator.
If you want this to compete with tools in the OpenClaw space, I’d prioritize first class Telegram and Slack support. Push progress into a chat thread, and let me approve, retry, cancel from there. That’s where teams live. A separate mobile frontend will always feel clunky and fragile.
Worth noting that this is currently broken for a number of users, I'm on a Max plan and I get the message "Error: Remote Control is not enabled for your account. Contact your administrator" which isn't helpful since I'm my administrator and ... this gets recursive quickly.
Can anyone recommend a tool that gives a 'mission control' overview of multiple agents, but also combines some basic project management functionality.
For example, maybe I have an idea for a feature and I want to spin up a new branch and have agents work on that. But then I get stuck or bored (I'm talking personal usage), so decide to park it. But maybe after a few days I have a shower thought and want to resume it.
The current method of listing sessions and resuming them can work, but you need to find the right session. If there is something that shows all the branches, a docs overview of what that feature it, and the current progress it would make this workflow a lot more effective. Plus I switch LLMs when I hit rate limits.
I'm probably going to just build it myself, but wondering if anyone has something that does this already.
Maybe combine Claude Code + Obsidian, so Claude can use the node structure as a second brain for projects. I was just watching this video (not affiliated):
Native macOS and iOS apps, backed by a rust binary that I put anywhere and connect to. Right now I’m just LAN but eventually will tailscale.
Works with claude and codex. Both passively watching an active CLI session for both and you can take over those sessions if needed and interact in the app
This new remote control handoff is neat but still requires you to remember to do the handoff. Oftentimes I’m waiting on an agent and then walk away.
I built Crabigator[1] and it's a wrapper around `claude` and `codex`, so its ready for coding on the go on start and already streaming. Plus, crabigator shows many parallel windows, separated by repo/project/machine, so you can manage multiple agents seamlessly.
I would rather have them commit to make a standard out of --sdk-url. I really want to use it in production, but it being undocumented means they can take it away anytime, so stdout it has to be (and hooks).
I don't think they target the pros here who already solved this problem with vpn/tmux/ssh but to those whose thrilled serious reaction will be "whoaaa crazy i can command claude code now from my phone while on the toilet or on a date?" It's basically a defense attempt against Openclaw.
I would have hoped for them to at least support the "/clear" command or some form of it, especially to manage context if we're limited to a single session between the terminal and Claude iOS app. I like to work on things one at a time and /clear my way between them to get back to 0% context, which seems impossible with the current setup here?
Typing "/clear" in the terminal clears it, but the Claude iOS app just outputs raw XML instead and doesn't actually do anything:
This kind of release shows Anthropic as a company is suffering from the same thing we all are right now. Removing the friction from having an idea and executing it stops you from remembering The Point. Yes, programming from your phone is an exciting modality and maybe even the future of how we work, but coding from your bedroom, AND the toilet, AND the woods AND your office is definitely (hopefully) not the future.
I wonder if is anyone working on an AI framework that encourages us to keep our eye on the big picture, then walk away when a reasonable amount of work is done for the day.
Yes, individuals are creating cool mobile coding solutions and Anthropic doesn't want to get left behind. I know I'm working my ass off at work right now because LLM coding makes it fun, but I also often don't prioritize what I'm doing for the big picture because I just try every thing that comes into my inbox, in order, because it's so fast to do with Claude Code.
There are two types of software engineers: Those who do and then think, or those who think and then do. Claude Code seems to strictly be for the former, while typically the engineers who can maintain software long-term are the latter.
Not sure if we have any LLM-tooling for the latter, seems to be more about how you use the tools we have available, but they're all pulling us to be "do first, think later" so unless you're careful, they'll just assume you want to do more and think less, hence all the vibeslop floating around.
> Claude Code seems to strictly be for the former, while typically the engineers who can maintain software long-term are the latter.
Given the number of CC users I know who spend significant time on creating/iterating designs and specs before moving to the coding phase, I can tell you, your assumption is wrong. Check how different people actually use it before projecting your views.
Yeah, I wasn't trying to say "These are the people who use CC, for these purposes" but rather what the intention seems to for Claude Code in the first place. I'm using CC from time to time, to keep up to date with what tooling is available, and also know people who use CC every day and plan a lot up front, sorry if I gave the impression that I meant that everyone using CC is doing that, was trying to get at what the purpose of the tool seems to be, which seems to be true today too, as the models continuously seem to steer you to "doing" and moving faster, not stopping and thinking.
This seems like a real coarse and not particularly accurate binary, but even if it were true, the thing about Claude Code and agentic coding like this is the cost of making a mistake or the cost of not being happy with a design and having to back it out is getting smaller and smaller.
I would argue that rapidly iterating reveals more about the problem, even for the most thoughtful of us. It's not like you check your own reasoning at the door when you just dive head first into something.
This isn't a binary thing - even if you prefer to build maintainable systems very often the trade-off is - you don't ship in time and there's no long term - the project gets scrapped.
So even if it comes at the expense of long term maintainability - everyone should have this in their toolbox.
I find it often helps me to see a feature before I evaluate if it was really a good idea in the first place. This is my failing--but one thing I like about Claude is that it's now possible to just try stuff and throw away whatever doesn't work out.
I usually have conversations with Claude for clearing my mind and forming the scope of a project. I usually use voice transcription from Claude app to take notes and explore all my options.
Same. When I can't be at my desk, my projects don't stop -- I just do the tasks that work well enough on the phone. Brainstorming, planning, etc. Or tasks that the agent can easily verify.
Having access to my local repository and my whole home folder is much easier than dealing with Claude or ChatGPT on the web. (Lots of manual markdown shuffling, passing in zipfiles of repositories, etc).
id say claude code is designed for think then do - thats where its different from other tools!
i think it still pulls to do then think because you cant tell what the agent understood of what you asked it to do from that first think, until its actually produced something.
I agree in your basic framing but not your conclusion. Met plenty of do-ers before thinkers that are self-aware enough to also maintain software longterm.
Claude Code and similar agents help me execute experiments, prototypes and full designs based on ideas that I have been refining in my head for years, but never had the time or resources to implement.
They also help get me past design paralysis driven by overthinking.
Perhaps the difference between acceleration and slop is the experience to know what to keep, what to throw away, and what to keep refining.
This is the real insight in this thread. The false binary of "rest OR work" is dissolving. I do some of my best problem-solving while walking my kid to school or making lunch...the context switch lets things percolate. Having a way to capture that momentum without needing to rush back to my desk and remember what I was thinking would be genuinely useful. The interface matters less than the latency between idea and execution.
How is this not solved by a simple voice recorder? You can process and act on it later while not forgetting your thoughts when inspiration hits. People have been doing that for at least like 50 years now.
"The false binary of "rest OR work" is dissolving."
If you're like most people in this forum, there are people who stand to gain financially if you convince yourself that you don't need boundaries between work and rest. You may even believe that you stand to gain financially, and that this will be best for you in the long term.
Please, take some time to rest for a day or two and really think about what you want your boundaries to be. Write them down.
> The false binary of "rest OR work" is dissolving
Sounds like someone hasn't yet worked multiple years with software engineering, or any job for that matter.
Your mind might trick you into believing it won't matter, but your body and mind NEEDS to be disconnected from work, 100%, at some point during your regular rhythms of life, otherwise you'll burn out much faster than the people you seemingly are trying to compete with.
Life never been a sprint, but it is a marathon, and if you spend all your young experience-less years on treating it as a sprint, you won't have any energy left for completing the marathon.
I’m guessing you’re suggesting it’s ok to lose time if you’re away from your computer enjoying life, and I agree. I also don’t see the issue in finding ways to be save time with work.
If you mean something different, please elaborate.
This seems like an excellent thread to plug the TUI I've been working on that makes using bubblewrap relatively easy and somewhat pleasant. I have a recipe in the README for using it with Claude. Granted that Claude has --sandbox, but probably better that sandboxing be done by something outside of the Anthropic ecosystem.
I think a significant distinction between your approach and Claude’s approach is that your approach requires allowing your machine to accept inbound connections but Claude’s approach does not. Claude probably went with the latter to avoid a whole class of security issues and mitigate risk of users having their machines compromised. I’m not familiar with what the new vectors of attack are with Claude’s approach though.
That's not what vendor lock in means. If you sign up for a cloud hoster and then build your whole product on propriety services that you can't get anywhere else instead of using an off the shelf database or open source software, that's vendor lock in.
If you'd have to switch to a different tool to do your coding that's not vendor lock in.
I’ve been doing this with a tmux tunnel and an app on my laptop that connects sessions you select to a virtual terminal using sockets. I asked Claude to build it and it works great - full terminal functionality and Markdown review with comments so you don’t need to cross your eyes to review plans.
Excited to see how this matures so people without that inclination can also be constantly pestered by the nagging idea that someone, somewhere is being more productive than them :)
except there will be no dropbox moment. There is no startup that stands a chance, Openclaw is free, the foundation model providers basically won this space just by providing subscriptions cheaper than any competitor could ever do.
One more step closer to a closed source system. I think their objective is to move all your code on their systems so you can only modify the code through their AI so they have a moat and will be difficult to move away. They will “guard” your source code and you’ll never see it.
Imagine if tomorrow they make a 10x smarter AI, but they say: you can only use if you upload your source code to us and you can’t see anymore the source code.
So you either stay on lower end models or you give up and use a 10x model.
I only see one issue: will be very difficult for them to “guard” the source code and don’t let you access.
Seems like it could be problematic in the future since code can't be fixed by humans so the only source of future code for training is unedited Claudeslop.
Boggles my mind that this is actually a thing that still needs to be solved. Just remote into your computer (I prefer TeamViewer). That is it. One step.
> Unlike Claude Code on the web, which runs on cloud infrastructure, Remote Control sessions run directly on your machine and interact with your local filesystem. The web and mobile interfaces are just a window into that local session.
For the vibe'y workflows, this would easily solve parallel long running work without skipping permissions: schedule 10 different tasks and go for a run. Occasionally review what the hallucination machine wants to do, smash yes a few times, occasionally tell it not to be silly, have a nice run. Essentially, solving remote development, though perhaps not quite in the way how people usually think of it.
> Limitations
> One remote session at a time: each Claude Code session supports one remote connection.
Ehh, I think it's hardly different from the people who leave Claude Code working on problems overnight with really loose permissions - seemingly the chance of them returning to it mining crypto for Putin is low enough for it to not be a consideration (see the whole OpenClaw movement).
And people have been remoting into their machines for a while, so now having a pretty-UI-but-walled-garden variety doesn't ring that many alarm bells. If they manage to get it right, it wouldn't be that much different from running some CI stuff on your machine while you're making tea, or reviewing pull requests while lounging around.
Small UX note: the first time you run the command it only shows a URL. It's not until you run it again that you discover it also generates a QR code, which is actually the fastest way to open it on your phone. Would be nice if the QR showed up on the first run too, almost missed it.
Does anyone know if it caffeinates automatically? I sometimes see caffeinate appear in the terminal tab title so clearly they are using it, but I’m just curious if I have to run caffeinate separately if, for instance, the agent finishes its task and is waiting for a new one and I want to keep it alive.
Regular claude code is already a remote access door to your setup, once you've granted a few command execution permissions. (e.g. if it can edit your code and run the test suite)
I was just thinking about it two days ago - how nice it would be to use my local Claude code instead of the limited cloud version to make some ad hoc changes when I have a fresh idea on a hike. And two days later - here we go, a new release
This resonates hard. I'm a self-taught dev who started coding ~7 months ago, and honestly the conversational back-and-forth with Claude is how I built my entire first app. Not by reading docs cover to cover, but by describing what I wanted, getting code back, breaking it, asking why, and iterating. The idea of doing that untethered from my desk is genuinely exciting — not because I want to work more, but because some of my best thinking happens on walks, not in front of a screen.
How about they are pointing out a worrisome direction society might be taking, whereas work will infiltrate even more what used to be family or personal time, thus accelerating burnout?
have you gotten a terminal interface on your phone to be acceptably usable? I haven't - not without a real keyboard attached in any case. too many parts of the UX are designed for a true keyboard.
Someone is going to solve this with a non-buggy app, but it really needs to have all the features of Claude code. Everyone is a power user in this segment
Oh come on, now that I have a personal remote control already set up using hooks, specifically the PermissionRequest, and Home Assistant push notifications where I can allow or deny a specific action?
TIL that HA notifications can have associated actions. I have the exact same setup as you, except I only receive the notification and then walk over to the laptop to unblock the agent feeling like a human tool call. This will improve my workflow, thank you.
The notification payload for reference, you will also need a permission input_select (pending/allow/deny) and an automation that triggers upon mobile_app_notification_action:
Exactly that. And the push notification includes what I am approving. Also with some sensible delay in sending out these pushes, because otherwise I may be bombarded with push notifications, while already having it manually approved.
Yep. Came to say the same thing. I'd only used Codex in VSCode and in the Codex app, and at least those have the same history, but my understanding is that the cloud and CLI versions have this hierarchy of 'visibility' [0]. Perhaps they'll need to change this design decision?
Anthropic is spitting out software in 2 weeks that took enterprises 24 months to ship 5 years ago (and was still buggy AF, let's please actually think about all the vmware citrix enterprise trash you tolerated). It'll get hardened over the next couple weeks.
You all can pretend the software dev cycle hasn't changed... get real.
I honestly think this is definitely where (at least part of) the industry is heading, yes.
This is not to say engineers are getting replaced — but, certainly, they are changing their work. And, sure, maybe _some_ of them are being replaced. Not most of the ones I know, though. They are essential to orchestrate, curate, maintain, and drive all of this.
(Now, do they want to orchestrate it? Whole different story...)
WOW I had been using the Codex app (Claude/Anthropic have a few annoying problems) and wishing there was something like this!
I often get ideas while I'm in bed or outside away from my computer, and was thinking that the ability to code on your computer from your phone, through AI, would be such a killer app.
My favorite use case would be asking the AI to review code and going over its findings/suggestions while I'm away from the computer or trying to fall asleep.
Doesn't have to be. Before OpenClaw was a thing, people were experimenting with setups to allow them to drive their agent remotely.
And of course, OpenClaw is built to be a very generalist agent with a chat interface - same effective outcome as remotely controlling an AI harness, but not exactly what everyone wants.
Pretty happy to see this. I've previously tried happy.engineer for this, but that wanted my Anthropic API token for itself (!) which is a no-no.
Seeing how the labs tend to copy the best functionality in any FOSS developments, I decided to wait - happy I did, here's the official functionality for this that is much more trustworthy.
This is an extremely clunky and buggy prerelease, so don't try to hot fix prod from the toilet without a different mobile frontend.
Right now:
- You can't interrupt Claude (you press stop and he keeps going!)
- At best it stops but just keeps spinning
- The UI disconnects intermittently
- It disconnects if you switch to other parts of Claude
- It can get stuck in plan mode
- Introspection is poor
- You see XML in the output instead of things like buttons
- One session at a time
- Sessions at times don't load
- Everytime you navigate away from Code you need to wait for your session to reappear
I'm sure I'm missing a few things.
Exactly my experience, I know they vibe code features and that’s fine but it looks like they don’t do proper testing which is surprising to me because all you need bunch of cheap interns to some decent enough testing
No there is a wide gap between good and bad testers. Great testers are worth their weight in gold and delight in ruining programmer's days all day long. IMO not a good place to skimp and a GREAT place to spend for talent.
Thats not true. Even for testing things, you need to do thoroughly now because standards are high.
From where I'm viewing, the standards in software have never been lower.
> all you need bunch of cheap interns to some decent enough testing
Sounds like a problem AI can easily solve!
Struggled with it also, given up (for now).
I thought coding was a solved problem Boris?
I have little doubt where things are going, but the irony of the way they communicate versus the quality of their actual product is palpable.
Claude Code (the product, not the underlying model) has been one of the buggiest, least polished products I have ever used. And it's not exactly rocket science to begin with. Maybe they should try writing slightly less than 100% of their code with AI?
More generally, Anthropic's reliability track record for a company which claims to have solved coding is astonishingly poor. Just look at their status page - https://status.claude.com/ - multiple severe incidents, every day. And that's to say nothing of the constant stream of bugs for simple behavior in the desktop app, Claude Code, their various IDE integrations, the tools they offer in the API, and so on.
Their models are so good that they make dealing with the rest all worth it. But if I were a non-research engineer at Anthropic, I wouldn't strut around gloating. I'd hide my head in a paper bag.
Even when it's operating normal the webapp is constantly crashing.
Mobile app stops working..
It's a pain.
At least right now.
I am constantly amazed how developers went hard for claude-code when there were and are so many better implementations of the same idea.
It's also a tool that has a ton of telemetry, doesn't take advantage of the OS sandbox, and has so many tiny little patch updates that my company has become overworked trying to manage this.
Its worst feature (to me at least), is the, "CLAUDE.md"s sprinkled all over, everywhere in our repository. It's impossible to know when or if one of them gets read, and what random stale effect, when it does decide to read it, has now been triggered. Yes, I know, I'm responsible for keeping them up to date and they should be part of any PR, but claude itself doesn't always even know it needs to update any of them, because it decided to ignore the parent CLAUDE.md file.
Coding is a solved problem. Problems with the code - these are far from solved, in fact they're multiplying, but coding is definitely solved
What does "solving" coding mean?
> What does "solving" coding mean?
Maybe this was sarcasm, but it's a good point:
"Coding" is solved in the same way that "writing English language" is solved by LLMs. Given ideas, AI can generate acceptable output. It's not writing the next "Ulysses," though, and it's definitely not coming up with authentically creative ideas.
But the days of needing to learn esoteric syntax in order to write code are probably numbered.
It types code, wallah!
It is solved in his org. He never promised quality software, though.
You get a buggy electron app and they get billions in valuation.
Clearly no one values quality anymore. 1000% yolo
He is trolling to increase the stock price before IPO
Generating code is a solved problem. Some people think that is the same thing.
Until problems are a solved problem, I feel I'm ok.
On top of that is something they should have had from earlier times. My biggest pain point is to not to be able to continue from my phone. I just use a service to pipe telegram to any cc session in the dev machine. This is the number 1 reason why I got excited by openclaw in the first place but its overkill to have it just to control cc
That's a bummer. I was looking forward to testing this, but that seems pretty limiting.
My current solution uses Tailscale with Termius on iOS. It's a pretty robust solution so far, except for the actual difficulty of reading/working on a mobile screen. But for the most part, input controls work.
My one gripe with Termius is that I can't put text directly into stdin using the default iOS voice-to-text feature baked into the keyboard.
I’ve been doing this for a while [1], but ultimately settled on a building a thin transport layer for Telegram to accept and return media, and persistent channels, vastly improved messaging UX, etc. and ended up turning this into a ‘claw with a heartbeat and SOUL [2].
[1] https://elliotbonneville.com/phone-to-mac-persistent-termina...
[2] https://elliotbonneville.com/claude-code-is-all-you-need/
I really enjoyed reading both posts. Thanks for sharing!
I, like many others, have written my own "claw" implementation, but it's stagnated a bit. I use it through Slack, but the idea of journaling with it is compelling. Especially when combined with the recent "two sentence" journaling article[1] that floated through HN not too long ago.
[1] https://alexanderbjoy.com/two-sentence-journal-approaches/
Happy you liked it! Always really nice to get positive feedback.
I’ll have to check out the journaling article. I’ve been journaling a lot more lately!
Great posts! So far [2] is the only "claw" that has caught my interest, mostly because it isn't trying to do everything itself in some bespoke, NIH way.
I've been using email and Cloudeflare email router. You don't get the direct feedback of a terminal, but it's much easier to read what's happening in html formatted email.
It also feels kind of nice to just fire off an email and let it do it's thing.
same but i use android. so i just talk to a google keep note and then copy/paste it. helpful for longer things
How can the like most popular terminal emulator not accept voice input? That's crazy why hasn't someone made something better?
Wispr Flow on mobile fills this gap.
it works in Blink. is there a better terminal in ios i should use
I use opencode web (server running on my desktop) and accessing it from my phone and it works well.
Same here. So I have to resort to speaking elsewhere (notes app) and copying/pasting.
Echo supports this.
Why couldn't they prompt Claude code to fix all the issues?
There are probably multiple Claude agents running as we speak trying to fix the issues.
Does that mean more issues will show up soon?
It's outsourced to Codex
You're not making either of them look good. Maybe they should have used Gemini?
They say a picture is worth a thousand words, they should've just used Nano Banana
I think they should be aware that CC is big enough codebase that they can't vibe code anymore.
Sounds like something that was vibe coded :)
I'm willing to bet most of their libraries are definitely vibe coded. I'm using the claude-agent-sdk and there are quite a few bugs and some weird design decisions. And looking through the actual python code it's definitely not what I would classify 'best practice'. Bunch of imports in functions, switching on strings instead of enums, etc.
I had to downgrade to an earlier release because an update introduced a regression where they weren't handling all of their own event types.
A few weeks ago the github integration was completely broken on the claude website for multiple days. It's very clear they vibe code everything and while it's laudable that they eat their own dogfood, it really projects a very amateurish image about their infrastructure and implementation quality.
Remember 100% of Claude Code is written by Claude
check out codecast.sh its so far ahead :)
> - You can't interrupt Claude (you press stop and he keeps going!)
This is normal behavior on desktop sometimes its in the middle of something? I also assume there's some latency
> - At best it stops but just keeps spinning
Latency issues then?
> - It can get stuck in plan mode
I've had this happen from the desktop, and using Claude Code from mobile before remote control, I assume this has nothing to do with remote control but a partial outage of sorts with Claude Code sometimes?
I don't work for Anthropic, just basing off my anecdotal experience.
Latency is like what, 50ms? You can’t explain these with latency. It’s just slop work from Claude.
oh. i was excited for a native alternative to happy coder, or sshing to a tmux session, but i guess not:/
Posted elsewhere but will copy here. Been doing this for a while.
- - -
get tailscale (free) and join on both devices
install tmux
get an ios/android terminal (echo / termius)
enable "remote login" if on mac (disable on public wifi)
mosh/ssh into computer
now you can do tmux then claude / codex / w/e on either device and reconnect freely via tmux ls and tmux attach -t <id>
- - -
You can name tmux and resume by name via tmux new -s <feature> and tmux attach -t <feature>
I have just today discovered zmx [1] which is like tmux but I always hated the tmux terminal emulation and how it hijacks scrolling, especially on Termius on my phone. It does session persistence but I think without the terminal emulator side of things, so scrolling works normally.
Been testing it today with Claude Code and it seems to work quite well switching between my laptop and phone.
[1] https://github.com/neurosnap/zmx
I think everyone started out with this
I think people familiar with tmux/screen and tailscale or how to setup their own VPN might have.
Based on my experience many people don't know this is a thing you can do.
I'm still doing this. Is there a better setup? It's already very convenient and secure.
Do you have an alias or something so that every time you open CC, tmux is running?
How do you deal with multiple concurrent sessions of CC with this setup?
How important is mosh? I wasn't able to get it set up the last time I tried... ran into a bunch of issues.
I always setup my terminals to open a tmux session by default (or attach to any existing one). For example, in my ~/.config/alacritty/alacritty.toml:
tmux supports tabs so you can have multiple Claude Code sessions running concurrently. You do need to learn a few tmux keyboard shortcuts to use it effectively (e.g. opening/closing/switching tab)."tmux ls" shows you all the open tmux sessions
Could even use cc to check in on and/or "send-keys"
What wasn't working about mosh? Just install mosh and use mosh to connect
I recommend Eternal Terminal instead of mosh for this.
Yes. Doing the same. What is the advantage of this new feature? Tmux/Tailscale/Termius give you full control of your terminal. Or mainly to save the end user the hassle to set it up correctly?
> Or mainly to save the end user the hassle to set it up correctly
It's this.
Don't have a Dropbox moment ;) [1]
[1]: https://news.ycombinator.com/item?id=9224
Oh lots of people will not be comfortable with tmux approach. The anthropic feature makes sense. But it's Max only and doesn't work well according to other comments.
What I posted "just works".
Ease of setup is the biggest reason. I use this setup as well, but there are other UX niceties that would be a lot better with a dedicated mobile app: push notifications when Claude needs your input (I use a hook for this that connects to Pushover, but that's another service and extra setup), voice input, autocorrect that's right for this context, etc.
I also do this!
I used it to add a MIDI driver and support to my OS this afternoon. Worked okay, but I agree it is a bit clunky yet. I think it is pretty good for a preview release. Much better than nothing.
Opencode's 'web' command makes your local session run on the browser with same access rights as the cli. It's a pretty slick interface too. I sometimes use it instead of the cli even when I can access both.
You can test it right now if you want with the included free models.
https://opencode.ai/docs/web/
I was having too many bugs using it with my phone, I gave up and went back to Termux
It's changing super fast. I am using it on the desktop mostly and when I tried on my phone there were issues yes. But do try it out again in a few weeks.
(I am actually using zellij on the remote and using various CLIs more than I am using only opencode on the web. I was using wezterm mux until about a week ago but the current state of the terminal is not very good for this scenario. It seems like almost all the CLIs are choking because of nodejs ink library)
There's https://happy.engineering/ which already does this with many fewer bugs and supports codex.
I was using this religiously but there’s a bug currently that makes the initialization fail and/or throws an error on the phone client. Absolutely great piece of software otherwise, free, anonymous, encrypted and so on. Really hope the team can fix this soon - I would hate to switch back to tmux tunneling.
I feel like a lot of folks are saying this kills the Code on your Phone opportunity some start-ups are building for. I don't agree. I feel like coding agents are like streaming services, we will subscribe to multiple and switch between them. So for one there's value in a universal control plane. The other is that mobile as a coding interface should offer more than a remote control to the desktop. I think there's still some space to cook, especially if people are investing 8 hours a day talking to agents, the interface surely matters.
I don't know a single person who is satisfied with the status quo on streaming services where you have to subscribe to multiple ones. Everyone is complaining that the landscape is 1) more fragmented than cable was, 2) costs more, 3) has even more ads than cable
I think people forgot how bad it was. It was much more fragmented before but instead of services it was fragmented by time. Sure you have access to Seinfeld, but you can watch one or two Seinfelds a night at 8pm and 11pm.
I also remember base cable without any movies was around $60 or something and with some movie channels is >$100. And that's not inflation adjusted. You can easily get 3 or 4 of the top services for $100 today.
Finally claiming there are more ads on these services is a joke. There was ~20m for every 30m of programming, meaning 1/3 of the time you're watching commercials. And not just any commercials, the same commercials over and over. There was even a case of shows being sped up on cable to show more commercials.
I get it, everyone wants everything seamlessly and for next to nothing, but claiming that 90s cable was even comparable is absurd.
https://www.digitaltrends.com/home-theater/how-networks-spee...
Not that it is particularly relevant to agentic coding but how can anyone truly argue streaming costs more? Average cable packages were exceeding 125-150 USD a month (in 2000 dollars). Under no circumstances would I be sympathetic to the argument that streaming costs more.
You can get all 7 of the major streaming subs for less without even shopping out deals. That is 100s of times the volume and quality of content that was delivered on cable for far less. It is so much content realistically that no one I have ever met has subscribed to all of them at once.
The argument really is empty. The fragmentized experience is annoying, but it isn't more expensive...And it DEFINITELY has fewer ads.
I'm in central europe, atm 70 TV channels is $15/month.
You can't seriously claim points 2) and 3) if you've ever actually paid for and watched cable
I'm in central europe, atm 70 TV channels is $15/month.
Its just amazing how people on HN can say the most absurd things with total conviction. No wonder LLMs do the same, it's in the training data.
I literally see no ads on my streaming subscription for close to a tenth of the price of cable.
you have just one streaming subscription?
I do. I rotate every few months among different services. I don't keep a single service permanently.
I think “streaming is more expensive now than cable before” crowd are people like my wife that have 8-10 of them
I'm using copilot on vscode and the agent is "Auto" which cost 10% less.
The results are enough for me and I'm not doing things that allow me to differentiate the output between ChatGPT, Claude and, the others.
The agents are more like the radio in my car, whenever I want music, I switch channel until I find something good enough.
If I'm really in need of something special, I'll use Spotify on my phone.
And sometimes, I just drive with the radio off.
I agree. I spend a lot of time working from my phone so I had to make my own workflow that works for me. I've been following all these bans and drama with the subscription keys and custom harnesses etc. I think there's room for a "universal control plan" that lets you leverage the CLI providers (and whatever crappy interfaces / apis they give you).
There's a comparison of the approaches as I see them here https://yepanywhere.com/subscription-access-approaches
Weird all these companies struggle so much to support remote services, ssh has been working for me pretty seamlessly for like the 20 years I've been using it and has allowed me to remote-control any computer I own with relatively reliable authentication (with some hiccups that tend to be patched pretty rapidly when found) throughout that entire period. I hear tell it worked even before I was using computers professionally, too
Yeah but this is from an AI company so its mad different
Yea I've noticed that most things made by a company are worse and less reliable than things that are free and maintained by volunteers
This has always been the case, for pretty much everything. There's a reason the world's infrastructure runs on FOSS.
This is buggy to no end. Anthropic needs to slow down
The daily “what broke and changed now” with claude code is wearing me out fast.
Claude Code is a good product, they should just keep on steadily improving it and improving the model. I am not sure why they are spraying in all directions like this..
Claude Code Team: Please fix the core experience instead of branching out into all these tertiary features. I know it is fun and profit to release new features but you need to go deeper into features not broader into there be dragons territory.
I find Xcode's Claude Agent client to be much more stable than the Claude web and Mac apps.
This is a classic case of vibe coding gets you 80% of the way there, but polishing takes the other 80%.
There is that (you can still use coding agents on the other 80% to polish, by the way) but to me this situation underscores the value of a good "editor"--someone who says this is good to ship vs. not this not now.
Maybe it’s related to what I tend to use the agents for, but I guess I don’t understand what is this for. Typically I try to structure the tasks in a way that require me to do or check something important when the agent gets back to me. If the agents query is trivial enough I can respond from my phone, it was likely not needed at all. If the agent finished - fine. It will have to wait until I get back in front of the computer anyway.
I've used similar things (omnara/happy) while taking walks. Sometimes I'll get an idea about the problem I'm working on and I can just dictate it into my phone and check in 15min later. I stopped being able to do that when claude added those nice interview panes to clarify things because it didn't work back then. But mostly it's really annoying when you think you've created the plan/prompt and that it's ready to go. But it gets stuck or decided to stop while you're away. I pretty often need to give Claude a "continue" kick. To be fair this happens far less after Opus 4.6.
Also, I felt the need to use it far more when I was on Pro vs a Max plan. On Pro when you hit the usage windows it's nice to be able to kick claude back into gear without scheduling your life around getting back to the terminal to type "continue".
- Plan mode -> answer questions/make corrections, continue planning
- Some of us don't do full yolo mode all the time, then tool approvals or code reviews are required, nice to do a quick review and decide if you need to go back to your computer or not
- Letting claude spin or handle a long-running task outside of normal work hours and being able to check in intermittently to see if something crashed
I don't dangerously accept permissions outside of a few scripts I have reviewed as safe. This means claude gets stuck often when testing it's work, but also means it doesn't uninstall production workloads from the kubernetes cluster.
We've re-invented GNU screen in the most inefficient way imaginable
Well it DOES have less storage than a Nomad (hence lame), but this way you don't need to pay for a public IP address, or for a VPS to run Wireguard on, or for a commercial VPN solution, and then install a terminal emulator on your phone and set up SSH keys.
People tried reinventing terminals, SSH, and tmux for phones. It's a pretty terrible experience using your thumbs. And it takes significant know-how to set up.
And in modern stacks, it almost necessitates a man in the middle - tailscale is common but it's still a central provider. So is it really the most inefficient way possible?
Fair point technically, but I think the value proposition isn't the persistent session, rathere it's the abstraction layer. Screen/tmux assumes you know what commands to run. This assumes you know what outcome you want. For someone like me who came to coding late and doesn't have 20 years of muscle memory with terminal tools, the inefficiency in transport is more than offset by the efficiency in intent. Different tools for different people.
tmux/screen is literally less work to use than this thing.
You need to learn to type less than a dozen total characters including the command.
Not to mention a lot of terminals automatically integrate with tmux so you don’t have to do anything but open the terminal.
Sure, different tools for different people. And if you want to use a new fangled triangular wheel they just invented, no one’s going to stop you
It’s still a triangular wheel at the end of the day
That’s not at all how this works. Commands are relayed through Anthropic’s servers with a client polling mechanism.
Right, that's the "most inefficient way possible" (though personally I disagree, there are more inefficient ways to be found).
You could put the transport protocol on the blockchain, I suppose
You are making me more creative.
we can upload snapshot of zip files to blockchain, then notify customer via servers
Even better, train an entirely new LLM with your prompt added to its data set. It will be imbued with its own latent sense of purpose. All you need to do after that is type "let there be light!"
I'm probably 10 years out of date. Are ethereum smart contracts still a thing? I'm sure you could deploy one of those for every agent session to handle the notifications
Yes, that's a significantly less efficient way to manage persistent sessions
I'm pretty sure "how do we disallow running our agents in screen sessions" is on a jira board at some places
I’m running the agent in tmux in a colo. When I’m at a computer I use that, when I’m on the go the RC app is more convenient
Too many limitations, for now I'll stick with self-hosted https://github.com/tiann/hapi and Tailscale
What are the drawbacks of HAPI? Seems too good to be true. Will give it a try.
Doesn't look like it has proper worktree management. UIs that abstract away worktrees are very powerful. I vibe coded my own (https://github.com/9cb14c1ec0/vibe-manager), which unfortunately doesn't have the remote component that hapi does.
My needs are very basic and it hasn’t failed me yet, I like that it doesn’t try to do much. I know it has voice capabilities through eleven labs but I haven’t used that feature.
I just built yoloAI https://github.com/kstenerud/yoloai
Gives me full sandboxing with bypass permissions, tmux, and cherry-pick level control over what gets pulled back out into my work dir.
Mix in tailscale and I can control it from anywhere, on any device, with full transferrability using established and battle proven tooling.
I've been running something similar for a few months, which is a voice-first interface for Claude Code running on a local Flask server. Instead of texting from my phone, I just talk to it. It spawns agents in tmux sessions, manages context with handoff notes between sessions, and has a card display for visual output.
The remote control feature is cool but the real unlock for me was voice. Typing on a phone is a terrible interface for coding conversations. Speaking is surprisingly natural for things like "check the test output" or "what did that agent do while I was away."
The tmux crowd in this thread is right that SSH + tmux gets you 90% of the way there. But adding voice on top changes the interaction model. You stop treating it like a terminal and start treating it like a collaborator.
Here is a demo of it controlling my smart lights: https://www.youtube.com/watch?v=HFmp9HFv50s
If you want this to compete with tools in the OpenClaw space, I’d prioritize first class Telegram and Slack support. Push progress into a chat thread, and let me approve, retry, cancel from there. That’s where teams live. A separate mobile frontend will always feel clunky and fragile.
Worth noting that this is currently broken for a number of users, I'm on a Max plan and I get the message "Error: Remote Control is not enabled for your account. Contact your administrator" which isn't helpful since I'm my administrator and ... this gets recursive quickly.
There's an open issue on github for it:
https://github.com/anthropics/claude-code/issues/28098
claude /logout -> claude /login -> claude /remote-control
If you'd read the entire issue, you'd see that not only is that solution mentioned multiple times, it's not working for some people.
it worked for me
If it worked for everybody that issue would already be closed.
On Android app it needs Claude GitHub connection with scope to act on my behalf! Otherwise it won't work in the app. Really do NOT like that!
Why does the remote control needs that? For what?
I rather use the common developer tools like termux or mosh etc. on a phone if I need that functionality.
You can scope it to repos. Make a repo just for what you want. That's your sandbox.
Make a throwaway GitHub account just for it and give it PR access to your private repos.
But the whole point of remote control was to avoid that situation.
that's actually a good idea. Thanks, was not thinking about such a workaround!
Can anyone recommend a tool that gives a 'mission control' overview of multiple agents, but also combines some basic project management functionality.
For example, maybe I have an idea for a feature and I want to spin up a new branch and have agents work on that. But then I get stuck or bored (I'm talking personal usage), so decide to park it. But maybe after a few days I have a shower thought and want to resume it.
The current method of listing sessions and resuming them can work, but you need to find the right session. If there is something that shows all the branches, a docs overview of what that feature it, and the current progress it would make this workflow a lot more effective. Plus I switch LLMs when I hit rate limits.
I'm probably going to just build it myself, but wondering if anyone has something that does this already.
Maybe combine Claude Code + Obsidian, so Claude can use the node structure as a second brain for projects. I was just watching this video (not affiliated):
https://youtu.be/6MBq1paspVU
Honestly, build it yourself. I think we’re firmly in the personal software era (or disposable).
But if you don’t want to, I’ve been building basically this https://github.com/Robdel12/OrbitDock
Native macOS and iOS apps, backed by a rust binary that I put anywhere and connect to. Right now I’m just LAN but eventually will tailscale.
Works with claude and codex. Both passively watching an active CLI session for both and you can take over those sessions if needed and interact in the app
Conductor (mac app) does some of this, might want to take a look.
New speed record for time from LaunchHN to getting Sherlocked set by that company everyone was dunking on a couple of weeks ago
This new remote control handoff is neat but still requires you to remember to do the handoff. Oftentimes I’m waiting on an agent and then walk away.
I built Crabigator[1] and it's a wrapper around `claude` and `codex`, so its ready for coding on the go on start and already streaming. Plus, crabigator shows many parallel windows, separated by repo/project/machine, so you can manage multiple agents seamlessly.
[1]: https://drinkcrabigator.com
I would rather have them commit to make a standard out of --sdk-url. I really want to use it in production, but it being undocumented means they can take it away anytime, so stdout it has to be (and hooks).
I don't think they target the pros here who already solved this problem with vpn/tmux/ssh but to those whose thrilled serious reaction will be "whoaaa crazy i can command claude code now from my phone while on the toilet or on a date?" It's basically a defense attempt against Openclaw.
I would have hoped for them to at least support the "/clear" command or some form of it, especially to manage context if we're limited to a single session between the terminal and Claude iOS app. I like to work on things one at a time and /clear my way between them to get back to 0% context, which seems impossible with the current setup here?
Typing "/clear" in the terminal clears it, but the Claude iOS app just outputs raw XML instead and doesn't actually do anything:
Running Claude Code from a phone just seems like a recipe for Alzheimer’s. Rest, then focus and build.
This kind of release shows Anthropic as a company is suffering from the same thing we all are right now. Removing the friction from having an idea and executing it stops you from remembering The Point. Yes, programming from your phone is an exciting modality and maybe even the future of how we work, but coding from your bedroom, AND the toilet, AND the woods AND your office is definitely (hopefully) not the future.
I wonder if is anyone working on an AI framework that encourages us to keep our eye on the big picture, then walk away when a reasonable amount of work is done for the day.
Yes, individuals are creating cool mobile coding solutions and Anthropic doesn't want to get left behind. I know I'm working my ass off at work right now because LLM coding makes it fun, but I also often don't prioritize what I'm doing for the big picture because I just try every thing that comes into my inbox, in order, because it's so fast to do with Claude Code.
We all sense it!: <https://newsroom.haas.berkeley.edu/ai-promised-to-free-up-wo...> <https://ghuntley.com/teleport/> <https://steve-yegge.medium.com/the-ai-vampire-eda6e4f07163>
There are two types of software engineers: Those who do and then think, or those who think and then do. Claude Code seems to strictly be for the former, while typically the engineers who can maintain software long-term are the latter.
Not sure if we have any LLM-tooling for the latter, seems to be more about how you use the tools we have available, but they're all pulling us to be "do first, think later" so unless you're careful, they'll just assume you want to do more and think less, hence all the vibeslop floating around.
> Claude Code seems to strictly be for the former, while typically the engineers who can maintain software long-term are the latter.
Given the number of CC users I know who spend significant time on creating/iterating designs and specs before moving to the coding phase, I can tell you, your assumption is wrong. Check how different people actually use it before projecting your views.
Yeah, I wasn't trying to say "These are the people who use CC, for these purposes" but rather what the intention seems to for Claude Code in the first place. I'm using CC from time to time, to keep up to date with what tooling is available, and also know people who use CC every day and plan a lot up front, sorry if I gave the impression that I meant that everyone using CC is doing that, was trying to get at what the purpose of the tool seems to be, which seems to be true today too, as the models continuously seem to steer you to "doing" and moving faster, not stopping and thinking.
This seems like a real coarse and not particularly accurate binary, but even if it were true, the thing about Claude Code and agentic coding like this is the cost of making a mistake or the cost of not being happy with a design and having to back it out is getting smaller and smaller.
I would argue that rapidly iterating reveals more about the problem, even for the most thoughtful of us. It's not like you check your own reasoning at the door when you just dive head first into something.
This isn't a binary thing - even if you prefer to build maintainable systems very often the trade-off is - you don't ship in time and there's no long term - the project gets scrapped.
So even if it comes at the expense of long term maintainability - everyone should have this in their toolbox.
I find it often helps me to see a feature before I evaluate if it was really a good idea in the first place. This is my failing--but one thing I like about Claude is that it's now possible to just try stuff and throw away whatever doesn't work out.
Was always possible. Now just easier.
I usually have conversations with Claude for clearing my mind and forming the scope of a project. I usually use voice transcription from Claude app to take notes and explore all my options.
Same. When I can't be at my desk, my projects don't stop -- I just do the tasks that work well enough on the phone. Brainstorming, planning, etc. Or tasks that the agent can easily verify.
Having access to my local repository and my whole home folder is much easier than dealing with Claude or ChatGPT on the web. (Lots of manual markdown shuffling, passing in zipfiles of repositories, etc).
id say claude code is designed for think then do - thats where its different from other tools!
i think it still pulls to do then think because you cant tell what the agent understood of what you asked it to do from that first think, until its actually produced something.
I agree in your basic framing but not your conclusion. Met plenty of do-ers before thinkers that are self-aware enough to also maintain software longterm.
I would definitely disagree.
Claude Code and similar agents help me execute experiments, prototypes and full designs based on ideas that I have been refining in my head for years, but never had the time or resources to implement.
They also help get me past design paralysis driven by overthinking.
Perhaps the difference between acceleration and slop is the experience to know what to keep, what to throw away, and what to keep refining.
These coding tools work better when you think and play first before doing...
One could just as easily argue hunching over your desk staring at your computer has neurological implications.
My favorite way to vibe code is by voice while in the hot tub. Rest AND focus AND build.
This is the real insight in this thread. The false binary of "rest OR work" is dissolving. I do some of my best problem-solving while walking my kid to school or making lunch...the context switch lets things percolate. Having a way to capture that momentum without needing to rush back to my desk and remember what I was thinking would be genuinely useful. The interface matters less than the latency between idea and execution.
How is this not solved by a simple voice recorder? You can process and act on it later while not forgetting your thoughts when inspiration hits. People have been doing that for at least like 50 years now.
"The false binary of "rest OR work" is dissolving."
If you're like most people in this forum, there are people who stand to gain financially if you convince yourself that you don't need boundaries between work and rest. You may even believe that you stand to gain financially, and that this will be best for you in the long term.
Please, take some time to rest for a day or two and really think about what you want your boundaries to be. Write them down.
> The false binary of "rest OR work" is dissolving
Sounds like someone hasn't yet worked multiple years with software engineering, or any job for that matter.
Your mind might trick you into believing it won't matter, but your body and mind NEEDS to be disconnected from work, 100%, at some point during your regular rhythms of life, otherwise you'll burn out much faster than the people you seemingly are trying to compete with.
Life never been a sprint, but it is a marathon, and if you spend all your young experience-less years on treating it as a sprint, you won't have any energy left for completing the marathon.
Take care of yourself, your mind and your body.
Wait why should I prefer being stuck in the office over taking a walk and periodically steering Claude code by phone?
On the other hand, you lose a lot of time if you step away from a session and it gets stuck asking for permission to do something simple.
Oh no! Anyway.
I don’t understand this comment?
I’m guessing you’re suggesting it’s ok to lose time if you’re away from your computer enjoying life, and I agree. I also don’t see the issue in finding ways to be save time with work.
If you mean something different, please elaborate.
Could I not achieve this with a ssh client on my android, and use regular Claude code remotely that way? Or am I missing something?
This seems like an excellent thread to plug the TUI I've been working on that makes using bubblewrap relatively easy and somewhat pleasant. I have a recipe in the README for using it with Claude. Granted that Claude has --sandbox, but probably better that sandboxing be done by something outside of the Anthropic ecosystem.
https://github.com/reubenfirmin/bubblewrap-tui
That's what I've been doing with termux, mosh, and tmux.
I think a significant distinction between your approach and Claude’s approach is that your approach requires allowing your machine to accept inbound connections but Claude’s approach does not. Claude probably went with the latter to avoid a whole class of security issues and mitigate risk of users having their machines compromised. I’m not familiar with what the new vectors of attack are with Claude’s approach though.
Yeah the remote control featureset is pretty limited right now. I did a comparison here https://yepanywhere.com/claude-code-remote-control/ (with my own project). I'm sure they'll iterate on it. Overall it's such an obvious feature for them to add I'm surprised it took them so long to ship. There are probably at least 50 such projects that people have made (https://github.com/kzahel/yepanywhere/blob/main/docs/competi...)
The one feature drawback of tailscale/tmux/termius is no file upload. And ergonomics, ability to view files/diffs easily, though that's subjective.
Perhaps it took a while to figure out how to do it over HTTP, especially the security stuff.
With e.g. tmux you'll piggyback on decades of SSH development.
> SSH development.
Or Mosh, just like OP said. Mosh handles interruptions much better than SSH does
As I understand it, Mosh piggybacks on SSH. Have they recently dropped the SSH negotiation?
Which is so much better because you can do other terminal stuff and you can avoid vendor lock in.
That's not what vendor lock in means. If you sign up for a cloud hoster and then build your whole product on propriety services that you can't get anywhere else instead of using an off the shelf database or open source software, that's vendor lock in.
If you'd have to switch to a different tool to do your coding that's not vendor lock in.
In this case you are locking your workflow to the vendor's solution.
I’ve been doing this with a tmux tunnel and an app on my laptop that connects sessions you select to a virtual terminal using sockets. I asked Claude to build it and it works great - full terminal functionality and Markdown review with comments so you don’t need to cross your eyes to review plans.
Excited to see how this matures so people without that inclination can also be constantly pestered by the nagging idea that someone, somewhere is being more productive than them :)
so is harnessing tmux/tailscale new "rsync/FTP is enough" thing nowadays?
Perhaps, but the key difference is that it’s a developer tool you use from the command line.
except there will be no dropbox moment. There is no startup that stands a chance, Openclaw is free, the foundation model providers basically won this space just by providing subscriptions cheaper than any competitor could ever do.
One more step closer to a closed source system. I think their objective is to move all your code on their systems so you can only modify the code through their AI so they have a moat and will be difficult to move away. They will “guard” your source code and you’ll never see it.
Imagine if tomorrow they make a 10x smarter AI, but they say: you can only use if you upload your source code to us and you can’t see anymore the source code.
So you either stay on lower end models or you give up and use a 10x model.
I only see one issue: will be very difficult for them to “guard” the source code and don’t let you access.
Yeah, the only thing keeping them from doing that is that the open weight free models are just as good.
I wonder if they would take away your ability to prompt, maybe only letting you run agentically.
Seems like it could be problematic in the future since code can't be fixed by humans so the only source of future code for training is unedited Claudeslop.
check out codecast.sh for cross agent experience with this and much more!
Very nice, maybe my days of wasting my work breaks on HN are coming to an end.
Boggles my mind that this is actually a thing that still needs to be solved. Just remote into your computer (I prefer TeamViewer). That is it. One step.
This just points out that Claude is not on your computer, only your interface to your Claude session is.
> Unlike Claude Code on the web, which runs on cloud infrastructure, Remote Control sessions run directly on your machine and interact with your local filesystem. The web and mobile interfaces are just a window into that local session.
For the vibe'y workflows, this would easily solve parallel long running work without skipping permissions: schedule 10 different tasks and go for a run. Occasionally review what the hallucination machine wants to do, smash yes a few times, occasionally tell it not to be silly, have a nice run. Essentially, solving remote development, though perhaps not quite in the way how people usually think of it.
> Limitations
> One remote session at a time: each Claude Code session supports one remote connection.
Hmm. Give it 1-12 months.
This comment sounds like the basis for a nice Black Mirror episode.
Ehh, I think it's hardly different from the people who leave Claude Code working on problems overnight with really loose permissions - seemingly the chance of them returning to it mining crypto for Putin is low enough for it to not be a consideration (see the whole OpenClaw movement).
And people have been remoting into their machines for a while, so now having a pretty-UI-but-walled-garden variety doesn't ring that many alarm bells. If they manage to get it right, it wouldn't be that much different from running some CI stuff on your machine while you're making tea, or reviewing pull requests while lounging around.
all i get is: "Error: Remote Control is not enabled for your account. Contact your administrator."
Small UX note: the first time you run the command it only shows a URL. It's not until you run it again that you discover it also generates a QR code, which is actually the fastest way to open it on your phone. Would be nice if the QR showed up on the first run too, almost missed it.
You can also just open the app on your phone and go to the sidebar and click on Code and then you'll see the session at the top of your session list.
Oh nice, didn't know that. Thanks for the tip!
SSH app on your phone + Tailscale is already a much better experience.
Does anyone know if it caffeinates automatically? I sometimes see caffeinate appear in the terminal tab title so clearly they are using it, but I’m just curious if I have to run caffeinate separately if, for instance, the agent finishes its task and is waiting for a new one and I want to keep it alive.
I have been, doing doomcoding with tmux and ssh and oddly enough it was better experience than this.
I really don't want to trust an AI company with a remote access door on my setup
Regular claude code is already a remote access door to your setup, once you've granted a few command execution permissions. (e.g. if it can edit your code and run the test suite)
Yes and no: I hope (not verified) that regular claude code client only sends requests, and doesn't open ports for remote access
I wouldn't expect Remote Control to open any ports either
There's nothing stopping CC from spinning up a local service or running terminal commands to open ports.
I don't get it, just ssh?
How does this handle deauthentication / logging out all sessions?
Claude Code only supports logging out the current session via /logout
There's no logout all sessions equivalent unlike the web UI.
I was just thinking about it two days ago - how nice it would be to use my local Claude code instead of the limited cloud version to make some ad hoc changes when I have a fresh idea on a hike. And two days later - here we go, a new release
I feel closer to realizing my dream of walking by a forest and whispering back and forth to an LLM to get shit done
This resonates hard. I'm a self-taught dev who started coding ~7 months ago, and honestly the conversational back-and-forth with Claude is how I built my entire first app. Not by reading docs cover to cover, but by describing what I wanted, getting code back, breaking it, asking why, and iterating. The idea of doing that untethered from my desk is genuinely exciting — not because I want to work more, but because some of my best thinking happens on walks, not in front of a screen.
Wispr Flow just got an Android release, so everything except it talking back to you is now doable
Hoping OpenAI/ Codex will launch this soon too.
I run Codex using terminus, iOS, and a VPS.
Hoping Cursor will also adopt.
Cursor already has agents accessible from web/mobile https://cursor.com/blog/agent-web
But can those web/mobile-accessible agents be on your own hardware, e.g. your desktop at home?
Those are cloud agents.
even more reasons to sandbox it to a container or vm
Is this compatible with Bedrock powered Claude Code setups?
News flash...now you can continue to work whilst brewing a coffee, walking the dog or taking a shit.
jfc no
Are you actually complaining about having an option available, if you want it?
How about they are pointing out a worrisome direction society might be taking, whereas work will infiltrate even more what used to be family or personal time, thus accelerating burnout?
I don’t know what open claw is really, but I’ve heard so much about it like everyone else. Is this anthropic’s answer to that?
I’m genuinely wondering the same thing, seems like at least a precursor.
Claude -browser /Remote-controll ??? Win?
have been using tmux and ssh-on-the-phone doing for forever, what's new?
have you gotten a terminal interface on your phone to be acceptably usable? I haven't - not without a real keyboard attached in any case. too many parts of the UX are designed for a true keyboard.
I’ve had decent luck with Termius because it gives you a row of keyboard shortcuts above the usual keyboard. Still cramped, but it works.
Tmux is annoying with a mobile keyboard, so I vibe coded a little mobile-friendly wrapper https://github.com/zakandrewking/pocketbot
Someone is going to solve this with a non-buggy app, but it really needs to have all the features of Claude code. Everyone is a power user in this segment
take a look at opencode, it doesn't even have to be a terminal anymore to command your terminal from whatever device you are using
I used connectbot on android and it worked fine for me, the new 'terminal' with debian also worked well
Oh come on, now that I have a personal remote control already set up using hooks, specifically the PermissionRequest, and Home Assistant push notifications where I can allow or deny a specific action?
TIL that HA notifications can have associated actions. I have the exact same setup as you, except I only receive the notification and then walk over to the laptop to unblock the agent feeling like a human tool call. This will improve my workflow, thank you.
The notification payload for reference, you will also need a permission input_select (pending/allow/deny) and an automation that triggers upon mobile_app_notification_action:
Actionable notifications are a bit cumbersome on iOS since you need to long-press the notification for actions, but it does work.I'm trying to understand the setup you have here.
So your hook -> HA -> push notification? And then you just tap to approve?
Exactly that. And the push notification includes what I am approving. Also with some sensible delay in sending out these pushes, because otherwise I may be bombarded with push notifications, while already having it manually approved.
I want this for Codex.
Yep. Came to say the same thing. I'd only used Codex in VSCode and in the Codex app, and at least those have the same history, but my understanding is that the cloud and CLI versions have this hierarchy of 'visibility' [0]. Perhaps they'll need to change this design decision?
[0] https://www.youtube.com/watch?v=cczkDMmmrEE
At this point if one lab comes up with a feature it’s a matter of time before another does the same!
Anthropic is spitting out software in 2 weeks that took enterprises 24 months to ship 5 years ago (and was still buggy AF, let's please actually think about all the vmware citrix enterprise trash you tolerated). It'll get hardened over the next couple weeks.
You all can pretend the software dev cycle hasn't changed... get real.
So Microsoft/Github copilot was ahead of its time with AI driven PRs?
I honestly think this is definitely where (at least part of) the industry is heading, yes.
This is not to say engineers are getting replaced — but, certainly, they are changing their work. And, sure, maybe _some_ of them are being replaced. Not most of the ones I know, though. They are essential to orchestrate, curate, maintain, and drive all of this.
(Now, do they want to orchestrate it? Whole different story...)
WOW I had been using the Codex app (Claude/Anthropic have a few annoying problems) and wishing there was something like this!
I often get ideas while I'm in bed or outside away from my computer, and was thinking that the ability to code on your computer from your phone, through AI, would be such a killer app.
My favorite use case would be asking the AI to review code and going over its findings/suggestions while I'm away from the computer or trying to fall asleep.
Would be great if it supported API keys. I’m getting by with slack threads of all things for work.
I guess this is Anthropic's early version of a "claw"?
Doesn't have to be. Before OpenClaw was a thing, people were experimenting with setups to allow them to drive their agent remotely.
And of course, OpenClaw is built to be a very generalist agent with a chat interface - same effective outcome as remotely controlling an AI harness, but not exactly what everyone wants.
Pretty happy to see this. I've previously tried happy.engineer for this, but that wanted my Anthropic API token for itself (!) which is a no-no.
Seeing how the labs tend to copy the best functionality in any FOSS developments, I decided to wait - happy I did, here's the official functionality for this that is much more trustworthy.