Cursor often gets into "dead loops" trying to fix code [1][2]. But, Cursor also seems to get out of dead loops when it adds `console.log` statements everywhere.
We thought: "What if Cursor could access Node.js at runtime?". That would save it from adding `console.log` everywhere, and still get out of dead loops.
We looked into it and got Cursor to debug Node.js on its own.
It's a prototype, but if you're interested in trying it out, we'd love some feedback!
[2]: "We've all had the issue. You're trying to build a complex project with your AI companion. It runs into a dead loop, coding in circles, making suggestions it already tried that didn't work." ~Source: https://www.reddit.com/r/ChatGPTCoding/comments/1gz8fxb/solu...
> Cursor often gets into "dead loops" trying to fix code [1][2]. But, Cursor also seems to get out of dead loops when it adds `console.log` statements everywhere.
Not affiliated with Cursor, but I think there is a free plan to try Cursor out. It's still a bit janky in my opinion, but it's an interesting new dev experience.
I’ve been using Cursor extensively on TS projects and noticed a big improvement in speed and code quality by enforcing clear guardrails (strict types, linting, formatting) combined with rigorous TDD. When I hit an error or exception, I explicitly tell the agent to pause and first figure out the best way to debug and quickly pinpoint the root cause. This approach generally works well, provided I’ve decomposed the problem enough that we’re debugging just one thing at a time (though that’s not always trivial).
I’m interested to try this out to see if or how much it speeds things up further.
> When I hit an error or exception, I explicitly tell the agent to pause and first figure out the best way to debug and quickly pinpoint the root cause.
Out of curiosity, how do you tell Cursor about the bug? Do you copy/paste console messages from the terminal into the prompt?
I used to run into dead loops until I changed the way I prompt cursor. Asking it to focus on 1 error at a time massively improves the success rate. Had an old project in node 10 project that hadn’t been touched since ~2017 and after several attempts with cursor. Asked if to fix 1 issue at a time until it moved from node 10>12>14…>22 and after about 15 minutes it succeeded. That included updating all the dependencies and removing deprecated/dead/dropped packages and getting them to latest and updating Vue from 1 to 3.5
In cursor you can whitelist commands. So I white list npm run. Then tell cursor to upgrade a dependency and run npm run build and if there’s errors. Fix the first error and repeat. Once the build is successful, run npm test and fix 1 test at a time.
Takes some learning but once you figure it out it’s insanely powerful.
Related but I decided to try out Claude Code and one of the first things I noticed is that it didn't seem to be checking for basic transpilation errors (this was in a Typescript project). I added the following bit to the CLAUDE.md file (which is the equivalent of the assistant's LLM guidelines):
- `npm run build` - Run TypeScript compilation + Vite build. You should periodically run this to check for errors.
And it made a pretty significant difference when "autotake suggestions" was on.
Oh interesting, thanks for sharing! First time I hear about the `CLAUDE.md` file. Catching transpilation errors seems like a great use of custom instructions.
Do you have any other custom instructions in there?
2. Initialize everything using a template scaffolding system (vite, nextjs, etc.)
3. Create a simple README.md which contains a list of the desired tech stack (for reinforcement purposes) and goal of the project.
4. Start up Claude
5. Use the `/init` command which creates the vanilla CLAUDE.md file
Interestingly, Claude captures all the NPM commands from the package json into the instructional file but the issue with this is that in my particular test case it kept trying to run `npm run dev` which spins up the app itself. That's... not ideal since I'm running it myself so it can hot-reload and I can see the changes applied immediately. I usually delete that line.
But I should note that I've never used any "agentic coding assistants" before (Aider.chat, Cursor, etc.) - so I'm sort of learning as I go.
One thing I'd be curious about is if there is a "clientside browser" MCP that Claude (or other assistants) can monitor - because ideally you'd really want Claude to be doing the following periodically:
- Taking a screenshot of the browser and using some multimodal system to sanity check
- Monitor the Dev Tools console for errors/warnings/etc.
- Read through the DOM to sanity check
So I have "npm run dev" already running and I could theoretically instruct Claude (or another system) to attach itself to the Chromium/Firefox browser window.
This MCP server seems to be focused on Node/serverside debugging.
Thanks for the detailed write-up, this is super useful! I really appreciate it.
> Claude captures all the NPM commands from the package json into the instructional file
Oh neat, I didn't know that. I'll try to use `/init` next time!
> it kept trying to run `npm run dev` which spins up the app itself. That's... not ideal since I'm running it myself so it can hot-reload
+1 I regularly notice that behaviour too. I'm running node in dev mode, then Cursor wants to run node, but the port is already in use.
> I'd be curious about is if there is a "clientside browser" MCP
Great point, I noticed this is a common feature request for front-end development. I haven't used it myself, but multiple people mentioned BrowserTools [1]. I think it can read the Dev Tools console, but I'm not sure about multimodal or access to the DOM.
Oh neat, thanks so much for sharing. I was getting MCP client errors when simply using `console.error`. Sending log message from the MCP server to the MCP client via the logging capability makes a ton of sense!
> LLMs can not only add tons of log statements fast but parse the large output from a program faster.
Absolutely valid point, LLMs probably have an easier time adding log statements to the source code and parsing output from the console.
The debugger connection is "better" in certain cases because:
1. It avoids unnecessary diffs related to log statements. Currently, LLMs seem good at adding code, but forget to or struggle to delete code.
2. It avoids making logs noisy, although, admittedly, this can be managed with appropriate severity levels.
3. It provides more precise control over the program execution flow by stepping into, stepping over, and stepping out of lines.
And even that acronym expansion doesn't help, IMHO, whereas "a standard that's a cross between Language Server Protocol and OpenAPI, targeting LLM consumers" (emphasis on the API not AI) is hopefully a little better -- depending on whether the reader is already familiar with the problems that LSP and OpenAPI are solving
As far as I know there's still missing the UDDI-ish "discoverability" piece, although maybe one of the upcoming MCP standard releases will include /.well-known/mcp-servers type thing, or a SRV record, or both
Me neither: MCP Servers are systems that provide context, tools, and prompts to AI clients. They can expose data sources like files, documents, databases, and APIs (https://mcp.so/)
Oh cool, thanks for sharing! I like how your VS Code extension also adds the usual debugging cues like the red dot, the yellow line, and the call stack.
I didn't build any of that! I'm just interfacing with vs code debugger. It comes at no additional cost - it also means you are guaranteed all the other features of the standard vs code debugger as well, and can interrupt / collaborate / modify the debugging process alongside the llm.
Unless I'm misunderstanding this - Cline can do this off the shelf (no need to install & subscribe to cursor), I'm surprised you can't do this in cursor actually?
It doesnt help with all death loops, but we were able to better guide LLMs towards the core of the problem. In some cases like RPCs, it helped LLM investigate state and fix the issue better.
There's a facinating number of posts mentioning MCP in the last week or two. Many of them from young accounts, or accounts with a post years ago, then nothing, and suddenly MCP.
MCP itself just seems to be JSON-RPC with a schema to be self-describing, so it's not really that interesting that it warrants one post per day. The web page touts it as the magic needed to ensure you don't have to write API adapters for your LLM uses. Well, except you need to write or use an MCP server to do the translation. So unless you're using one that already exists (though hundreds have been advertised on HN recently,) you're just shifting the problem by adding a layer of indirection.
They also specifically say that authentication is something we'll work out later, so it seems none of these will have access to the data that actually has value.
I have no idea about this account specifically, but I'm starting to suspect there's a bot ring somewhere.
To be fair, after seeing the posts I installed the filesystem MCP server made by Anthropic and it is a massive game changer for working with Claude Desktop to program. I guided Claude to develop a 3000 line python wrapper for ffmpeg for my library conversion project as handbrake wasn't automatic enough, Tdarr was overkill for my 1 machine solution, and I wanted a clean interface as I'm not an expert in video encoding.
It has been possibly the most fun I've ever had programming, as it's been unbelievably effective and quick. I believe the hype. I'll note I'm also very wary of the potential of malicious MCP servers and am rather hesitant to use any "community" projects, as much as it saddens me to feel I'm losing my trust in the open source community.
I had a similar experience with the read-only Postgres MCP. Just asked Claude "how can I optimise the performance of my database?" and boom, it checked missing indexes, most used tables, work_mem, etc and gave personalized suggestions that made sense. It even used data provided by the `pg_stat_statements` extension automatically.
As simple as MCP conceptually can be, nothing like that was this simple and efficient before MCPs. It's a truly game-changer.
Interesting. Do you feel like creating an MCP for a niche devtool would lead you to use it or explore it more? Or is the wariness such that you wouldn't want to?
Personally I would hesitate to install any MCP server unless from an org I trust as it's a whole lot of extra code to trust/audit for possibly marginal gain. That said, depends what the niche is and if the use-case suits LLMs well.
Co-author here. We are not bots. Just two indie developers who quit their jobs trying to build impactful things.
'Vibe coding' has been great for us to prototype quickly, launch and test. As we kept building the complexity grew so much that we are hitting debugging death loops. Thats when we built this MCP.
My perspective is that MCPs are like browser extensions. There is nonpoint in building another IDE from scratch. IDEs like Cursor unlocked massive distribution channel and MCPs are hreat way to solve specific problems.
As you rightly observed, it's not the tech that is interesting. It's the ability of the tech to meet where users are is what makes it great. From my point of view, there is a specific context in which people vibe code - hobby projects, indie projects to make money, projects on legacy code, etc. They either might not have time or skill to do even simple things like adding an API. And thats where MCPs flourish. And as to why MCPs are making to the top, I think its because there are many of us who are matching the persona I mentioned.
MCP is a handy way to announce tools to the LLM it can use to interface with a service, usually remote. I use local mcp servers to pull on (authed) slack, notion, grafana and a few more into an agent like Cline. I don't think the HN post frequency is some bot thing, just seems to be something that's sticking with people as more and more mcp servers pop up. Makes wiring stuff together for agents to use easy to reason about.
I think I know of some of the other MCP posts you’re referring to. I saw multiple posts recently about AI controlling browsers using MCP and thought that was interesting. I’m not sure if it will be a fad or not, but I’ve been trying to model browsers as an HTTP resource instead as a side project. This allows you to HTTP DELETE a tab or POST /browsers to create a new browser. I think it might be a more natural way of using an API plus it can just use the classic auth in the form of API keys.
Its popular now because its showing up in tools people use so more backends are releasing MCP servers to access them, most are free so doubt its more than just a new fancy tool that people are excited about.
As for authentication most are locally run, or run on generic data like running a nodejs server mcp would be used locally by a team not exposed on the internet so access restrictions arent really needed
This seems like a good thread to post — can someone give an intuitive explanation of how MCP is supposed to work? I tried to set it up with LM Studio and a local LLM but couldn't for the life of me figure out what I actually need to do. MCP servers seems straightforward, I guess I need an MCP client somewhere, but I'm lost on how the LLM actually knows what tools are available and how/when a translation between "get x from external tool y" translates to actually getting `x` and how the LLM can then use it.
From what I could find online, it just works in Claude desktop app and there are some online efforts for mcp clients, but even ollama maintainers are confused about the implementation (https://github.com/ollama/ollama/issues/7865)
MCP really shines with a tool like Cline, Roo Code, or Cursor’s agent mode where an agent is writing code for you and needs access to tools. Some I’ve used successfully are Sentry for fixing errors or Figma for implementing a design. Most MCP servers are something you set up and run locally. It lets you set up auth once across tasks and configure which actions/tools are auto-approved so your agent can have higher autonomy during tasks. If you haven’t used Cursor or Cline you should give them a try.
Thank you for hints but the answer is about the same as I can find online - it just works with existing tools like Claude app or Cursor or Cline, but I specifically want to understand the mechanisms used and how I can take advantage of MCP servers using _local_ llms.
What’s wrong with the official MCP website? It guides you through building a server and an LLM-powered client. You can just have your local LLM operate the client and you’re set.
> “can you find all the tickets related to profile editing and combine them into an epic?”
That's a great use case. I imagine that makes you more productive because you can get multiple related issues out of the way by spending time in the same corner of your code base.
What other MCPs or use cases have caught your attention ?
Yeah I always use the analogy of “when I’m in the neighborhood, I’ll take care of that too” for related tickets.
I’ve had it update my tickets (like, leave a comment summarizing the code changes, transition the ticket.)
And I’ve described our workflow in a cursor rules file. And it knows to reassign the ticket to a certain person when I ask it to mark something ready for test. Or if I want to transition to a workflow strep 3 steps down, it knows my workflow from the rules file ands can perform each transition in sequence.
I believe that an authoritative implementation of an MCP server was just created in the last couple weeks, which explains the flurry of initial posts.
The value is that this interface is the other end of a socket; implement it, and your LLMs will have magic tools. Sure it’s just another JSON-RPC adapter, but it’s a JSON-RPC adapter that works and has working integrations.
Give it a try, if you remain open minded to the potential I suspect you can find a use case and application that would shock you in its efficacy.
You're surprised that people are talking about a new protocol from one of the biggest players in the most hyped space? Everyone is using MCP because it's becoming the industry standard.
You insinuating that Anthropic is using bot rings to promote MCP is not only nonsense, it's also breaking the HN guidelines.
Cursor often gets into "dead loops" trying to fix code [1][2]. But, Cursor also seems to get out of dead loops when it adds `console.log` statements everywhere.
We thought: "What if Cursor could access Node.js at runtime?". That would save it from adding `console.log` everywhere, and still get out of dead loops.
We looked into it and got Cursor to debug Node.js on its own.
It's a prototype, but if you're interested in trying it out, we'd love some feedback!
Github: github.com/hyperdrive-eng/mcp-nodejs-debugger
---
References:
[1]: "At this point I feel like giving up on Cursor and just accept that WE'RE NOT THERE YET." ~Source: https://forum.cursor.com/t/cursor-for-complex-projects/38911
[2]: "We've all had the issue. You're trying to build a complex project with your AI companion. It runs into a dead loop, coding in circles, making suggestions it already tried that didn't work." ~Source: https://www.reddit.com/r/ChatGPTCoding/comments/1gz8fxb/solu...
> Cursor often gets into "dead loops" trying to fix code [1][2]. But, Cursor also seems to get out of dead loops when it adds `console.log` statements everywhere.
Hah. That sounds like just... debugging.
Yes! Cursor really wants to debug. Personally, I don't love when Cursor adds 100s of `console.log` diffs to my project. I like debuggers for that.
Cursor does not want anything. Eseesntially it adds tons of console.log because it's the thing training material says it should do next.
That's what console ninja's new thing is supposed to do. I haven't used it yet as you have to pay.
Not affiliated with Cursor, but I think there is a free plan to try Cursor out. It's still a bit janky in my opinion, but it's an interesting new dev experience.
I haven't seen console ninja yet, thanks for mentioning it. It’s amazing!
Oh good point, I didn't realise Console Ninja [1] is a product. Thanks for sharing!
[1]: https://console-ninja.com
I’ve been using Cursor extensively on TS projects and noticed a big improvement in speed and code quality by enforcing clear guardrails (strict types, linting, formatting) combined with rigorous TDD. When I hit an error or exception, I explicitly tell the agent to pause and first figure out the best way to debug and quickly pinpoint the root cause. This approach generally works well, provided I’ve decomposed the problem enough that we’re debugging just one thing at a time (though that’s not always trivial).
I’m interested to try this out to see if or how much it speeds things up further.
Co-author here. We found it pretty useful in our personal projects. Would love to hear your feedback on how it went.
If you come across any issues specific to your stack, happy to look into them.
Thanks for sharing!
> When I hit an error or exception, I explicitly tell the agent to pause and first figure out the best way to debug and quickly pinpoint the root cause.
Out of curiosity, how do you tell Cursor about the bug? Do you copy/paste console messages from the terminal into the prompt?
I used to run into dead loops until I changed the way I prompt cursor. Asking it to focus on 1 error at a time massively improves the success rate. Had an old project in node 10 project that hadn’t been touched since ~2017 and after several attempts with cursor. Asked if to fix 1 issue at a time until it moved from node 10>12>14…>22 and after about 15 minutes it succeeded. That included updating all the dependencies and removing deprecated/dead/dropped packages and getting them to latest and updating Vue from 1 to 3.5
That's a great tip. Similarly, when implementing features, breaking down into simpler things and making it focus on each step works really well.
We (co-author here) even tried to build a design doc tool that can fed to Cursor in that effect.[1]
[1] https://updates.priyank.ch/projects/2024/03/01/introducing-a...
Thanks for sharing!
> Asking it to focus on 1 error at a time massively improves the success rate.
Out of curiosity, how do you tell Cursor about the bug? Do you copy/paste console messages from the terminal into the prompt?
In cursor you can whitelist commands. So I white list npm run. Then tell cursor to upgrade a dependency and run npm run build and if there’s errors. Fix the first error and repeat. Once the build is successful, run npm test and fix 1 test at a time.
Takes some learning but once you figure it out it’s insanely powerful.
Related but I decided to try out Claude Code and one of the first things I noticed is that it didn't seem to be checking for basic transpilation errors (this was in a Typescript project). I added the following bit to the CLAUDE.md file (which is the equivalent of the assistant's LLM guidelines):
And it made a pretty significant difference when "autotake suggestions" was on.Cursor does have a loop with Claude which fixes lint & type issues which is nice
Oh interesting, thanks for sharing! First time I hear about the `CLAUDE.md` file. Catching transpilation errors seems like a great use of custom instructions.
Do you have any other custom instructions in there?
So my basic approach is:
1. Create the project folder
2. Initialize everything using a template scaffolding system (vite, nextjs, etc.)
3. Create a simple README.md which contains a list of the desired tech stack (for reinforcement purposes) and goal of the project.
4. Start up Claude
5. Use the `/init` command which creates the vanilla CLAUDE.md file
Interestingly, Claude captures all the NPM commands from the package json into the instructional file but the issue with this is that in my particular test case it kept trying to run `npm run dev` which spins up the app itself. That's... not ideal since I'm running it myself so it can hot-reload and I can see the changes applied immediately. I usually delete that line.
But I should note that I've never used any "agentic coding assistants" before (Aider.chat, Cursor, etc.) - so I'm sort of learning as I go.
One thing I'd be curious about is if there is a "clientside browser" MCP that Claude (or other assistants) can monitor - because ideally you'd really want Claude to be doing the following periodically:
- Taking a screenshot of the browser and using some multimodal system to sanity check
- Monitor the Dev Tools console for errors/warnings/etc.
- Read through the DOM to sanity check
So I have "npm run dev" already running and I could theoretically instruct Claude (or another system) to attach itself to the Chromium/Firefox browser window.
This MCP server seems to be focused on Node/serverside debugging.
Thanks for the detailed write-up, this is super useful! I really appreciate it.
> Claude captures all the NPM commands from the package json into the instructional file
Oh neat, I didn't know that. I'll try to use `/init` next time!
> it kept trying to run `npm run dev` which spins up the app itself. That's... not ideal since I'm running it myself so it can hot-reload
+1 I regularly notice that behaviour too. I'm running node in dev mode, then Cursor wants to run node, but the port is already in use.
> I'd be curious about is if there is a "clientside browser" MCP
Great point, I noticed this is a common feature request for front-end development. I haven't used it myself, but multiple people mentioned BrowserTools [1]. I think it can read the Dev Tools console, but I'm not sure about multimodal or access to the DOM.
[1]: https://github.com/AgentDeskAI/browser-tools-mcp
Up to you, but there's actually a formal logging pathway[1] in MCP that has the potential to put this message somewhere that either of the two consumers[2] of your service could see it https://github.com/hyperdrive-eng/mcp-nodejs-debugger/blob/v...
1: https://spec.modelcontextprotocol.io/specification/2024-11-0...
2: the person, and (while I don't have evidence of it) I also can't think why the LLM couldn't see the message, either
Oh neat, thanks so much for sharing. I was getting MCP client errors when simply using `console.error`. Sending log message from the MCP server to the MCP client via the logging capability makes a ton of sense!
We'll have to look with fresh eyes on the print debugging vs debugger debate for LLMs.
LLMs can not only add tons of log statements fast but parse the large output from a program faster.
How exactly is debugger connection better for LLMs compared to log statements?
> LLMs can not only add tons of log statements fast but parse the large output from a program faster.
Absolutely valid point, LLMs probably have an easier time adding log statements to the source code and parsing output from the console.
The debugger connection is "better" in certain cases because: 1. It avoids unnecessary diffs related to log statements. Currently, LLMs seem good at adding code, but forget to or struggle to delete code. 2. It avoids making logs noisy, although, admittedly, this can be managed with appropriate severity levels. 3. It provides more precise control over the program execution flow by stepping into, stepping over, and stepping out of lines.
MCP = Model Context Protocol
(I didn't know what MCP was.)
And even that acronym expansion doesn't help, IMHO, whereas "a standard that's a cross between Language Server Protocol and OpenAPI, targeting LLM consumers" (emphasis on the API not AI) is hopefully a little better -- depending on whether the reader is already familiar with the problems that LSP and OpenAPI are solving
As far as I know there's still missing the UDDI-ish "discoverability" piece, although maybe one of the upcoming MCP standard releases will include /.well-known/mcp-servers type thing, or a SRV record, or both
UDDI = Universal Description, Discovery and Integration
(I didn't know what UDDI was.)
I instantly thought Master Computer Program, ngl.
I suspect that was intentional.
Me neither: MCP Servers are systems that provide context, tools, and prompts to AI clients. They can expose data sources like files, documents, databases, and APIs (https://mcp.so/)
Exciting to see more development in this direction!
I've been using a VS Code extension that I built to enable Claude / Cursor / Continue to automatically debug.
It's language agnostic and also uses MCP.
https://github.com/jasonjmcghee/claude-debugs-for-you
LSP and debugging support are incredibly useful for LLMs, and I'm hoping to see general support for these capabilities in the existing tools soon!
Oh cool, thanks for sharing! I like how your VS Code extension also adds the usual debugging cues like the red dot, the yellow line, and the call stack.
I didn't build any of that! I'm just interfacing with vs code debugger. It comes at no additional cost - it also means you are guaranteed all the other features of the standard vs code debugger as well, and can interrupt / collaborate / modify the debugging process alongside the llm.
Unless I'm misunderstanding this - Cline can do this off the shelf (no need to install & subscribe to cursor), I'm surprised you can't do this in cursor actually?
Good point, I haven't used Cline yet. It looks like it can access console logs [1]
[1]: https://docs.cline.bot/exploring-clines-tools/cline-tools-gu...This is really useful. LLMs need external feedback to self-correct reasoning.
https://arxiv.org/pdf/2310.01798
this doesnt really help get out of loops in my experience
none of these agents or ide can fix the underlying issue with the blackbox
until the blackbox is capable of genuine reasoning and awareness
calculated brute force is our only path forward
It doesnt help with all death loops, but we were able to better guide LLMs towards the core of the problem. In some cases like RPCs, it helped LLM investigate state and fix the issue better.
Does it work with bun?
Oh great question, we should look into other runtimes like bun, deno, etc.
Would you like to use it in a bun project?
Yes please :)
There's a facinating number of posts mentioning MCP in the last week or two. Many of them from young accounts, or accounts with a post years ago, then nothing, and suddenly MCP.
MCP itself just seems to be JSON-RPC with a schema to be self-describing, so it's not really that interesting that it warrants one post per day. The web page touts it as the magic needed to ensure you don't have to write API adapters for your LLM uses. Well, except you need to write or use an MCP server to do the translation. So unless you're using one that already exists (though hundreds have been advertised on HN recently,) you're just shifting the problem by adding a layer of indirection.
They also specifically say that authentication is something we'll work out later, so it seems none of these will have access to the data that actually has value.
I have no idea about this account specifically, but I'm starting to suspect there's a bot ring somewhere.
To be fair, after seeing the posts I installed the filesystem MCP server made by Anthropic and it is a massive game changer for working with Claude Desktop to program. I guided Claude to develop a 3000 line python wrapper for ffmpeg for my library conversion project as handbrake wasn't automatic enough, Tdarr was overkill for my 1 machine solution, and I wanted a clean interface as I'm not an expert in video encoding.
It has been possibly the most fun I've ever had programming, as it's been unbelievably effective and quick. I believe the hype. I'll note I'm also very wary of the potential of malicious MCP servers and am rather hesitant to use any "community" projects, as much as it saddens me to feel I'm losing my trust in the open source community.
I had a similar experience with the read-only Postgres MCP. Just asked Claude "how can I optimise the performance of my database?" and boom, it checked missing indexes, most used tables, work_mem, etc and gave personalized suggestions that made sense. It even used data provided by the `pg_stat_statements` extension automatically.
As simple as MCP conceptually can be, nothing like that was this simple and efficient before MCPs. It's a truly game-changer.
Check out also https://github.com/xataio/agent which builds on this idea in the form of a dedicated agent for Postgres.
Disclaimer: I’ve worked on this project.
The biggest game-changer is that whatever MCP server I need, I can just ask Claude to code it.
agreed if openai doesn't give their custom GPTs access to local stuff or implement MCP I might switch to claude soon
Interesting. Do you feel like creating an MCP for a niche devtool would lead you to use it or explore it more? Or is the wariness such that you wouldn't want to?
Personally I would hesitate to install any MCP server unless from an org I trust as it's a whole lot of extra code to trust/audit for possibly marginal gain. That said, depends what the niche is and if the use-case suits LLMs well.
I haven't used Claude yet. Can I ask much it cost to write those 3000 lines?
the irony is that ffmpeg-python exists (I'm using it in a bunch of scripts) and it also looks like much less than 3000 lines
It's fine for a library conversion project but I won't be surprised this is how things are headed in real production software too
[flagged]
You're off by three orders of magnitude. 3k lines. You also don't know what the wrapper does.
Co-author here. We are not bots. Just two indie developers who quit their jobs trying to build impactful things.
'Vibe coding' has been great for us to prototype quickly, launch and test. As we kept building the complexity grew so much that we are hitting debugging death loops. Thats when we built this MCP.
My perspective is that MCPs are like browser extensions. There is nonpoint in building another IDE from scratch. IDEs like Cursor unlocked massive distribution channel and MCPs are hreat way to solve specific problems.
As you rightly observed, it's not the tech that is interesting. It's the ability of the tech to meet where users are is what makes it great. From my point of view, there is a specific context in which people vibe code - hobby projects, indie projects to make money, projects on legacy code, etc. They either might not have time or skill to do even simple things like adding an API. And thats where MCPs flourish. And as to why MCPs are making to the top, I think its because there are many of us who are matching the persona I mentioned.
[flagged]
MCP is a handy way to announce tools to the LLM it can use to interface with a service, usually remote. I use local mcp servers to pull on (authed) slack, notion, grafana and a few more into an agent like Cline. I don't think the HN post frequency is some bot thing, just seems to be something that's sticking with people as more and more mcp servers pop up. Makes wiring stuff together for agents to use easy to reason about.
I've been aware of MCP for a long time but recently I started hearing about it a lot more, even from customers - it is clearly a hype.
A long time - are you confusing it with something else? It was only published 3 months ago, which would be why it has recently become a thing.
https://www.anthropic.com/news/model-context-protocol
How are you defining hype, here?
It's like saying that package manager for language X is a hype.
I think I know of some of the other MCP posts you’re referring to. I saw multiple posts recently about AI controlling browsers using MCP and thought that was interesting. I’m not sure if it will be a fad or not, but I’ve been trying to model browsers as an HTTP resource instead as a side project. This allows you to HTTP DELETE a tab or POST /browsers to create a new browser. I think it might be a more natural way of using an API plus it can just use the classic auth in the form of API keys.
relevant: TabFS <https://news.ycombinator.com/item?id=34847611> <https://github.com/osnr/TabFS> (MIT) and <https://github.com/balta2ar/brotab> (MIT) also showed up while I was trying to remember the name of TabFS
In case you haven't seen it, the MCP repo actually has Puppeteer as one of their example servers: https://github.com/modelcontextprotocol/servers/tree/2025.3....
Thank you! I had seen TabFS before, but the others are new to me. Will take a look! It's nice to see demand for this.
Its popular now because its showing up in tools people use so more backends are releasing MCP servers to access them, most are free so doubt its more than just a new fancy tool that people are excited about.
As for authentication most are locally run, or run on generic data like running a nodejs server mcp would be used locally by a team not exposed on the internet so access restrictions arent really needed
This seems like a good thread to post — can someone give an intuitive explanation of how MCP is supposed to work? I tried to set it up with LM Studio and a local LLM but couldn't for the life of me figure out what I actually need to do. MCP servers seems straightforward, I guess I need an MCP client somewhere, but I'm lost on how the LLM actually knows what tools are available and how/when a translation between "get x from external tool y" translates to actually getting `x` and how the LLM can then use it.
From what I could find online, it just works in Claude desktop app and there are some online efforts for mcp clients, but even ollama maintainers are confused about the implementation (https://github.com/ollama/ollama/issues/7865)
MCP really shines with a tool like Cline, Roo Code, or Cursor’s agent mode where an agent is writing code for you and needs access to tools. Some I’ve used successfully are Sentry for fixing errors or Figma for implementing a design. Most MCP servers are something you set up and run locally. It lets you set up auth once across tasks and configure which actions/tools are auto-approved so your agent can have higher autonomy during tasks. If you haven’t used Cursor or Cline you should give them a try.
Thank you for hints but the answer is about the same as I can find online - it just works with existing tools like Claude app or Cursor or Cline, but I specifically want to understand the mechanisms used and how I can take advantage of MCP servers using _local_ llms.
What’s wrong with the official MCP website? It guides you through building a server and an LLM-powered client. You can just have your local LLM operate the client and you’re set.
https://modelcontextprotocol.io/quickstart/server
https://modelcontextprotocol.io/quickstart/client
I’ve started using MCP just days ago and I feel the way I felt the first week I used ChatGPT.
It’s incredibly exciting. I started using the Atlassian MCP to get Claude to help me clean up our tickets.
Do things like “can you find all the tickets related to profile editing and combine them into an epic?”
> “can you find all the tickets related to profile editing and combine them into an epic?”
That's a great use case. I imagine that makes you more productive because you can get multiple related issues out of the way by spending time in the same corner of your code base.
What other MCPs or use cases have caught your attention ?
Yeah I always use the analogy of “when I’m in the neighborhood, I’ll take care of that too” for related tickets.
I’ve had it update my tickets (like, leave a comment summarizing the code changes, transition the ticket.)
And I’ve described our workflow in a cursor rules file. And it knows to reassign the ticket to a certain person when I ask it to mark something ready for test. Or if I want to transition to a workflow strep 3 steps down, it knows my workflow from the rules file ands can perform each transition in sequence.
I believe that an authoritative implementation of an MCP server was just created in the last couple weeks, which explains the flurry of initial posts.
The value is that this interface is the other end of a socket; implement it, and your LLMs will have magic tools. Sure it’s just another JSON-RPC adapter, but it’s a JSON-RPC adapter that works and has working integrations.
Give it a try, if you remain open minded to the potential I suspect you can find a use case and application that would shock you in its efficacy.
OP here, long-time lurker, first-time poster.
I agree with you, I also saw quite a few MCP posts on Show HN recently.
I didn't understand what the whole MCP thing was about, so my buddy and I hacked on this idea out of curiosity.
I almost didn't post it on Show HN, I thought "Ah no, nobody cares. It'll just disappear among the other posts".
Woke up Sunday morning, discovered the post is on the homepage of HN.
Re JSON-RPC, I don't have a lot to add to the conversation. I'll look into the MCP JSON-RPC spec though!
You're surprised that people are talking about a new protocol from one of the biggest players in the most hyped space? Everyone is using MCP because it's becoming the industry standard.
You insinuating that Anthropic is using bot rings to promote MCP is not only nonsense, it's also breaking the HN guidelines.
[flagged]
[flagged]