Trying to wrap my head around how it differs from my current on-the-go setup that is claude code. On claude's phone app or web app I can choose a repo, ask it for a feature and it writes the code, runs my tests + add more tests and creates a new branch. Then I can click on create PR or configure claude code to auto create PR.
I also have another setup which is a self-hosted docker compose behind my vpn with one container with claude code agents managed using agent of empires[0] and another container with playwright with sse. Using this setup, my agents get access to actual browser where it can test things live and I can access the app started by agent on domain:port. This is something I don't get with Claude.
[0] https://github.com/agent-of-empires/agent-of-empires
That setup is pretty much what we're trying to offer with Hoplite!
Using us means losing freedom and control with regards to infrastructure, however we think that's a tradeoff people would want to make in exchange for easier onboarding and a more polished experience.
Makes sense. It took me some experimentations with a few open source libraries and docker-compose to arrive at my setup. My setup also requires me to access agents via a terminal app on my phone so it's nice to have a web app like your offering.
Are you offering browser access to the agents in your setup?
Yes, agents have a persistent Chromium session they use via the agent-browser CLI. Typical workflow would involve starting the preview, seeding data, then the agent going through the old and new UX flows for a before + after view. We've also got some optimisations around saving the aforementioned flows in a QA library, so that they can be replayed without needing an agent to run through it all again.
>> It opens a pull request
>> Then keeps iterating as review comments land, in the same thread, with the same context.
My experience, particularly with Sol is that agents are generally really bad at knowing when to stop 'iterating' and will continue covering off ever-more obscure edge cases. How does Hoplite solve for this?
We've observed this pattern as well, and have counteracted it by keeping the tasks very finely scoped. For example, "The test(api) check has failed. Fix it, then immediately commit and push." -- or -- "The following comments have been added by reviewers. Resolve each one, then immediately commit and push."
This specificity helps Sol stay on track (most of the time). It doesn't work as well when the comment questions a complex piece of the architecture though.
Really interesting. I've been working on building out more self-sufficient agents with harnesses to match, and this feels like the next logical step. I'd love to try this.
Curious how you guys are thinking about the difference between project-specific and project-agnostic harnesses and tooling. For me, it feels like a lot of the work is project specific and I'm not sure how to abstract that.
In general, we've found that over the past couple of years agent harnesses have gotten much less restrictive, allowing the agent freedom to choose its own way of doing things. It seems like the project agnostic vs. specific harnesses will follow that trend. The pattern will work well for the current gen models, but eventually Fable 7 will be able to intuit how it should approach a specific project very well, at which point the challenge is making sure it has the tools to do so.
And I've made some changes to the coupon, does it work now?
Took me a minute to see the value - my first thought was "this is just cursor's cloud agents..."
But the key thing here for me is "Every sandbox boots your app on a live URL." Cursor doesn't easily have that, and that's what would allow me to ditch my local env entirely - the ability to actually try out a PR without needing to check it out locally.
So on that note: how does that work? We've had trouble with getting our dev env running in other cloud envs because it requires a few things (clickhouse, localstack, pg, etc) running which we manage via docker compose locally.
Also, some minor pricing feedback: it'd be really great if there were a version with pay-as-you-go and a cheaper fixed cost. I think your 99/seat/mo model is fine for professional work, but it's a lot to commit to for personal work.
Modal just released some features that would allow users to bring custom Docker images, and I'm working on getting your exact use case supported! Aiming to get it out by the end of the week.
Noted the pricing feedback! We're still figuring out exactly what works best so it's still very much so subject to change.
Looks neat! I've been using Amp (https://ampcode.com/) for a while (Which seems to be doing something very similar), and I really appreciate this type of workflow. One thread = one VM feels like a solid model going forward.
Agreed. Per-thread VMs are quite similar to how local agents use worktrees to avoid cross contamination, but with the added benefit of being able to easily scale up/down compute requirements on demand.
What’s the experience like going from an active on-going thread to a cloud-hosted one? I dont wanna always work on the cloud, and want my current setup to be exactly the same as in the cloud, and should be pretty seamless. Only folks i’ve seen solve these are folks who run the sandboxes locally and take that to the cloud like smolvm/microsandbox.
This is still something we're working on making seamless. The current approach is to install our MCP server and ask the local agent to start up a new thread on Hoplite when you want to transition to the cloud, but it doesn't carry over file system changes. (Unless you first push the contents to a remote branch, at which point the Hoplite agent can pull it down.)
just a data point, at my company we are building this internally. if you're targeting people building from scratch it might work, but there's no way you can port any somewhat mature infra stack, nor will the org want to. you'll need to deal with the variable complexity of everyone's dev environment which already doesn't work locally for thousands of different reasons.
Agreed that at the moment it's a very difficult problem, but one we're looking to solve! I think it becomes a no-brainer for most people if we're able to give each agent a replica of their production stack.
What does your current setup look like? And are you using an open source solution like OpenInspect for your in-house version, or building it from the ground up?
Building from ground up using OpenAI Agents SDK. We already have custom in house Cloud Development Environment, so the effort is just to "agent-ize" though which is not small. That's why I feel any team with any sort of infra support likely won't buy your product, since they already have the tribal knowledge to set this up. Newer teams might. But overall porting people's dev envs into the cloud is a tarpit problem(IMO), I was interested in this space too but decided against it for that reason. Happy to be proven wrong though!
I think for use cases like that, we'd offer on-prem deployments (similar to Factory), potentially coupled with a FDE. Still need to do a lot more research into the enterprise space.
I've been wondering what the alternatives to things like Github Copilot Cloud and Codex Cloud might be, especially ones that might be flexible wrt models, and this seems at least to have some of those behaviours.
If that perception is correct, please would you explain what it offers against those sorts of services (those in particular) and how the pricing compares - eg. their base levels are $20 a month, yours starts at a higher level - I can see there seems to be more brought in from the local IDE world (and similar), which seems very useful compared to the standard "prompt against repo, repeat" of the normal cloud agents but it would be useful to understand the targets and intents.
Thank you! Regarding models, as you said we're not locked into a specific provider, and are able to offer open weight models like Kimi K3 and GLM 5.2
Our pricing is higher than other providers because we do not upcharge on token or sandbox costs. We believe that people should be running as many agents as they possibly can handle, and an upcharge would create a monetary incentive for us to say that, when it's a genuine belief we hold.
We also offer features out of the box that would usually be behind enterprise gating (e.g. sandbox baking).
Amazing work, I am currently in the process of building something similar on my server for personal use, but yours looks really promising. Especially running sandboxes on your own can be tedious. Why have you opted for Modal instead of Firecracker or a similar micro VM solution?
Modal has a lot of small niceties that made them easy to implement, such as filesystem snapshots and programmatic build images. But I did see that AWS recently launched Lambda MicroVMs, and since we're an AWS house we may transition to using them.
My main issue with Modal is that their autoscaling is not as good as Daytona's. You have to stop the machine, resize, then start it, which takes ~3s and terminates all running processes. Daytona supports scaling up (but not down) without stopping the VM.
Also would recommend checking out ColeMurray/background-agents if you're planning to self host. Very good alternative! And the team behind it are great
You can do either. If you don't want to migrate over fully, I'd recommend setting up an automation to fix Sentry/PostHog issues as they come in. You can get a good feel for the platform and how it fits into your workflows that way.
We also have an MCP server that you can use to delegate tasks (e.g. research, debugging, SRE work) to Hoplite via your existing local setup.
exe.dev works quite well for giving an agent a computer and managing it remotely, but seems to require a fair bit more configuration to achieve parity with what we offer out of the box. Namely automations, PR autofix, visual QA, and general UI/UX polish.
I think it comes down to whether configuration or ease of use is valued more, and Hoplite favours the latter a bit more. (They shouldn't really be mutually exclusive, but we have a long way to go before we're happy claiming that we match/beat self-hosting in that area)
At the moment we're using Daytona as a redundant fallback in case Modal experiences an outage, but they have very stringent limits on how many resources we can consume concurrently. We're evaluating adding a second provider to help ease this so would love a chat! Feel free to email bence [at] hoplite.sh
Almost every possible name that is not a portmanteau, made-up word, or combination of N words has a naming "conflict". This is not an interesting thing to say and I wish that people would stop saying this on everything anyone ever posts on HN. It almost feels like it should be against HN rules to point out that something else shares the same name with no additional statements.
For one, it's interesting to me because I've already known Hoplite for years as nothing to do with AI. For two, I'm not sure how sharing that is so egregious it should be against the rules? Is there an interpretation I'm missing of my original comment? Does pointing out another Hoplite get interpreted as disparaging or accusatory in some way? Does it go against intellectual curiosity?
The term "conflict" implies "there is a problem here". I would say it's very lightly disparaging, because it implies the author didn't even do basic research on other things that are also named the thing they decided to call it.
I actually checked HN rules and just saw this:
> Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting.
Suggestion: showing an actual screenshot or video of your app is a much better indicator of effort than a generic Claude made animation. I've seen AI slop landing pages on far too many YC-backed startups. Not saying yours is one, but parts of it smell.
Agreed, we're working with a designer and are going to be fixing this very soon. Our main focus has been on making sure the product itself looks and feels very good to use -- probably not the best approach from a marketing POV.
Trying to wrap my head around how it differs from my current on-the-go setup that is claude code. On claude's phone app or web app I can choose a repo, ask it for a feature and it writes the code, runs my tests + add more tests and creates a new branch. Then I can click on create PR or configure claude code to auto create PR.
I also have another setup which is a self-hosted docker compose behind my vpn with one container with claude code agents managed using agent of empires[0] and another container with playwright with sse. Using this setup, my agents get access to actual browser where it can test things live and I can access the app started by agent on domain:port. This is something I don't get with Claude. [0] https://github.com/agent-of-empires/agent-of-empires
That setup is pretty much what we're trying to offer with Hoplite!
Using us means losing freedom and control with regards to infrastructure, however we think that's a tradeoff people would want to make in exchange for easier onboarding and a more polished experience.
Makes sense. It took me some experimentations with a few open source libraries and docker-compose to arrive at my setup. My setup also requires me to access agents via a terminal app on my phone so it's nice to have a web app like your offering.
Are you offering browser access to the agents in your setup?
Yes, agents have a persistent Chromium session they use via the agent-browser CLI. Typical workflow would involve starting the preview, seeding data, then the agent going through the old and new UX flows for a before + after view. We've also got some optimisations around saving the aforementioned flows in a QA library, so that they can be replayed without needing an agent to run through it all again.
>> It opens a pull request >> Then keeps iterating as review comments land, in the same thread, with the same context.
My experience, particularly with Sol is that agents are generally really bad at knowing when to stop 'iterating' and will continue covering off ever-more obscure edge cases. How does Hoplite solve for this?
We've observed this pattern as well, and have counteracted it by keeping the tasks very finely scoped. For example, "The test(api) check has failed. Fix it, then immediately commit and push." -- or -- "The following comments have been added by reviewers. Resolve each one, then immediately commit and push."
This specificity helps Sol stay on track (most of the time). It doesn't work as well when the comment questions a complex piece of the architecture though.
Really interesting. I've been working on building out more self-sufficient agents with harnesses to match, and this feels like the next logical step. I'd love to try this.
Curious how you guys are thinking about the difference between project-specific and project-agnostic harnesses and tooling. For me, it feels like a lot of the work is project specific and I'm not sure how to abstract that.
P.S. The code is not working for me.
In general, we've found that over the past couple of years agent harnesses have gotten much less restrictive, allowing the agent freedom to choose its own way of doing things. It seems like the project agnostic vs. specific harnesses will follow that trend. The pattern will work well for the current gen models, but eventually Fable 7 will be able to intuit how it should approach a specific project very well, at which point the challenge is making sure it has the tools to do so.
And I've made some changes to the coupon, does it work now?
Took me a minute to see the value - my first thought was "this is just cursor's cloud agents..."
But the key thing here for me is "Every sandbox boots your app on a live URL." Cursor doesn't easily have that, and that's what would allow me to ditch my local env entirely - the ability to actually try out a PR without needing to check it out locally.
So on that note: how does that work? We've had trouble with getting our dev env running in other cloud envs because it requires a few things (clickhouse, localstack, pg, etc) running which we manage via docker compose locally.
Also, some minor pricing feedback: it'd be really great if there were a version with pay-as-you-go and a cheaper fixed cost. I think your 99/seat/mo model is fine for professional work, but it's a lot to commit to for personal work.
Modal just released some features that would allow users to bring custom Docker images, and I'm working on getting your exact use case supported! Aiming to get it out by the end of the week.
Noted the pricing feedback! We're still figuring out exactly what works best so it's still very much so subject to change.
Looks neat! I've been using Amp (https://ampcode.com/) for a while (Which seems to be doing something very similar), and I really appreciate this type of workflow. One thread = one VM feels like a solid model going forward.
Agreed. Per-thread VMs are quite similar to how local agents use worktrees to avoid cross contamination, but with the added benefit of being able to easily scale up/down compute requirements on demand.
What’s the experience like going from an active on-going thread to a cloud-hosted one? I dont wanna always work on the cloud, and want my current setup to be exactly the same as in the cloud, and should be pretty seamless. Only folks i’ve seen solve these are folks who run the sandboxes locally and take that to the cloud like smolvm/microsandbox.
This is still something we're working on making seamless. The current approach is to install our MCP server and ask the local agent to start up a new thread on Hoplite when you want to transition to the cloud, but it doesn't carry over file system changes. (Unless you first push the contents to a remote branch, at which point the Hoplite agent can pull it down.)
just a data point, at my company we are building this internally. if you're targeting people building from scratch it might work, but there's no way you can port any somewhat mature infra stack, nor will the org want to. you'll need to deal with the variable complexity of everyone's dev environment which already doesn't work locally for thousands of different reasons.
Agreed that at the moment it's a very difficult problem, but one we're looking to solve! I think it becomes a no-brainer for most people if we're able to give each agent a replica of their production stack.
What does your current setup look like? And are you using an open source solution like OpenInspect for your in-house version, or building it from the ground up?
Building from ground up using OpenAI Agents SDK. We already have custom in house Cloud Development Environment, so the effort is just to "agent-ize" though which is not small. That's why I feel any team with any sort of infra support likely won't buy your product, since they already have the tribal knowledge to set this up. Newer teams might. But overall porting people's dev envs into the cloud is a tarpit problem(IMO), I was interested in this space too but decided against it for that reason. Happy to be proven wrong though!
I think for use cases like that, we'd offer on-prem deployments (similar to Factory), potentially coupled with a FDE. Still need to do a lot more research into the enterprise space.
Firstly: good luck!
I've been wondering what the alternatives to things like Github Copilot Cloud and Codex Cloud might be, especially ones that might be flexible wrt models, and this seems at least to have some of those behaviours.
If that perception is correct, please would you explain what it offers against those sorts of services (those in particular) and how the pricing compares - eg. their base levels are $20 a month, yours starts at a higher level - I can see there seems to be more brought in from the local IDE world (and similar), which seems very useful compared to the standard "prompt against repo, repeat" of the normal cloud agents but it would be useful to understand the targets and intents.
Thank you! Regarding models, as you said we're not locked into a specific provider, and are able to offer open weight models like Kimi K3 and GLM 5.2
Our pricing is higher than other providers because we do not upcharge on token or sandbox costs. We believe that people should be running as many agents as they possibly can handle, and an upcharge would create a monetary incentive for us to say that, when it's a genuine belief we hold.
We also offer features out of the box that would usually be behind enterprise gating (e.g. sandbox baking).
Amazing work, I am currently in the process of building something similar on my server for personal use, but yours looks really promising. Especially running sandboxes on your own can be tedious. Why have you opted for Modal instead of Firecracker or a similar micro VM solution?
Best of luck to you!
Modal has a lot of small niceties that made them easy to implement, such as filesystem snapshots and programmatic build images. But I did see that AWS recently launched Lambda MicroVMs, and since we're an AWS house we may transition to using them.
My main issue with Modal is that their autoscaling is not as good as Daytona's. You have to stop the machine, resize, then start it, which takes ~3s and terminates all running processes. Daytona supports scaling up (but not down) without stopping the VM.
Also would recommend checking out ColeMurray/background-agents if you're planning to self host. Very good alternative! And the team behind it are great
Thanks Bence, for your thoughts there. Will definitely take a look.
Looks interesting. Does this mean I would use this as my day to day harness? Or is it something additional to an established workflow?
You can do either. If you don't want to migrate over fully, I'd recommend setting up an automation to fix Sentry/PostHog issues as they come in. You can get a good feel for the platform and how it fits into your workflows that way.
We also have an MCP server that you can use to delegate tasks (e.g. research, debugging, SRE work) to Hoplite via your existing local setup.
Why would I use this over exe.dev?
exe.dev works quite well for giving an agent a computer and managing it remotely, but seems to require a fair bit more configuration to achieve parity with what we offer out of the box. Namely automations, PR autofix, visual QA, and general UI/UX polish.
I think it comes down to whether configuration or ease of use is valued more, and Hoplite favours the latter a bit more. (They shouldn't really be mutually exclusive, but we have a long way to go before we're happy claiming that we match/beat self-hosting in that area)
If you ever need to switch sandboxes, would be happy to chat.
At the moment we're using Daytona as a redundant fallback in case Modal experiences an outage, but they have very stringent limits on how many resources we can consume concurrently. We're evaluating adding a second provider to help ease this so would love a chat! Feel free to email bence [at] hoplite.sh
This Daytona? https://github.com/daytonaio/daytona
As the repo says no longer maintained
This one: https://www.daytona.io. Their platform was OSS for a long time but they decided to go closed source recently.
Naming conflict with https://www.hoplite.gg
Yeah, luckily they're in quite a different domain to us -- hopefully shouldn't have too much trouble winning the SEO battle
I don't imagine they get too many Google searches anyway, but they are incredibly popular in competitive Minecraft spheres.
Good luck, `hoplite` triggers a ton of Minecraft media for me and that name has been around 10 years.
Almost every possible name that is not a portmanteau, made-up word, or combination of N words has a naming "conflict". This is not an interesting thing to say and I wish that people would stop saying this on everything anyone ever posts on HN. It almost feels like it should be against HN rules to point out that something else shares the same name with no additional statements.
For one, it's interesting to me because I've already known Hoplite for years as nothing to do with AI. For two, I'm not sure how sharing that is so egregious it should be against the rules? Is there an interpretation I'm missing of my original comment? Does pointing out another Hoplite get interpreted as disparaging or accusatory in some way? Does it go against intellectual curiosity?
The term "conflict" implies "there is a problem here". I would say it's very lightly disparaging, because it implies the author didn't even do basic research on other things that are also named the thing they decided to call it.
I actually checked HN rules and just saw this:
> Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting.
So, there you go, I guess.
Suggestion: showing an actual screenshot or video of your app is a much better indicator of effort than a generic Claude made animation. I've seen AI slop landing pages on far too many YC-backed startups. Not saying yours is one, but parts of it smell.
Agreed 100%. We've been working with a designer on a complete redesign of our landing page to avoid that vibey-smell.
Your website text reads like Claude slop
Agreed, we're working with a designer and are going to be fixing this very soon. Our main focus has been on making sure the product itself looks and feels very good to use -- probably not the best approach from a marketing POV.