Ha! I love this. There's nothing like a proper Bash script to make me realize how terribly gross all of mine are.
The drum I'm currently beating is that local MCP is a ton of fun for techies like us - if you're on this website you can `npx ...` or install whatever you want with a modicum of common sense - but local MCP is going to be a dead end for mass adoption. If we want to build MCP servers that get used by everyday people (or on mobile or other locked down ecosystems) then remote MCP + OAuth is the only realistic way forward. I can't get my dad to open up a terminal window - anything over stdio or touching environment variables and API keys is a nonstarter.
Oh absolutely - but the infrastructure required to support a "click link, get remote MCP URL added to config automatically" flow is _so_ much smaller than the infrastructure required for a "click link, download and install arbitrary software that may or may not depend on having existing tools installed" flow.
Very cool! The docs here are a great overview of how MCP works, and a reminder to me of an old lesson:
We never should have abandoned REST. The whole point was for an interface to be self-describing; we wouldn't need MCP (or Swagger, or OpenAPI, etc) if we just stuck to REST instead of diverting down the JSON RPC route we've been on for 20 years.
And in what way is OpenAPI an abandonment of REST? It's an API documentation system that can be leveraged for generating REST server boilerplate code. If anything, it builds up the quality-of-life around REST.
I haven't seen a REST API in production for many years, maybe 15?
That's anecdotal obviously, but almost every, if not every, API I use today is an RPC call returning JSON.
Edit: to be clear, the distinction between what REST was defined as and what we use today often doesn't really matter. We use JSON APIs today, it is what it is. This is a case where it really matters though, LLM companies are now trying to push an entirely new protocol that tries to do roughly what REST did in the first place.
So the things we call "REST" in 2025 are not quite the same as the original specification of REST. One key aspect that has been abandoned is that sent data should be self-describing. That is, it shouldn't require any additional information to be useful. i.e. API documentation for JSON endpoints.
There's a great chapter on this in Hypermedia Systems[1]. Talks about both this and HATEOAS(Hypermedia as the engine of application state).
I have to say this is a very readable implementation to see how it all works in practice as well as a good reminder that it's a pretty simple universal tool interface.
>> a good reminder that it's a pretty simple universal tool interface.
That's because it's not really doing anything new. MCP is a land-grab by one company, quickly supported by the rest as they desperately work to abstract and supplant with their own "protocols". Welcome to the era of thin veneers that add little but complexity over what we already had.
Land grab? Not sure I'd call it that, but maybe it is.
I looked at it more as there were N different ways of configuring tools to be called, so they created N+1 (https://xkcd.com/927/) but have had good success at getting alignment on it.
While not knowing the reasons that OpenAI supported it, I would imagine it was along the lines of: "This is not more bad than the others we have seen, may as well use it"
Since that time, thankfully many folks have jumped into looking at it and making it better. I just wish the spec was good and easy to follow (I read through it, and I'm still looking for the real spec)
Not really in "pure bash". Also this needs to be labeled as a "toy".
Using an external tool like 'jq' especially written in C for parsing JSON, one can craft a exploitable JSON input to achieve code execution on the MCP server.
What could possibly go wrong? Maybe this CVE-2025-48060 [0] [1]?
Raw protocol, really. No marshaling, no conversions, none of the overhead from type management you get with modern Python, none of the turtles-all-the-way-down dependencies of NodeJS equivalents. I like it, although I would probably port it back to “lightweight” Python in about half the size :)
Some of us still prize compute efficiency, especially those who have been using Python for a long time and are contemplating the new kinds of code patterns that have emerged from data science...
I don't think they are comparable. MCPShell is a go program to run shell scripts, while the other one allows to define MCP operations as bash functions.
Not quite the same. The bash sdk can't be used to run arbitrary shell commands any more than to run arbitrary python programs.
Based on your recent comment history vibes, I'm pretty sure that you're using a LLM to post comments; and the remainder is self-promotion towards your own articles. Please don't do that here; HN aims to have interesting and not generic conversations: https://news.ycombinator.com/newsguidelines.html
fwiw I don't have a problem with LLM posts. But I do agree that this is pretty generic. If you want to use an LLM to post comments: find a better prompt / workflow
LLM-generated comments are not wanted on HN; we want to preserve it as a place for discussion between humans.
Also, comments of the format "here's what an LLM said about this topic:" are best avoided. We don't want to normalise a style of discussing issues in which we generate an LLM output and make that the central conversation topic; we prefer original human thought here.
I guess if we're gonna do monkey's paw/work to rule type of interpretations, HN should just add "we value authentic human experiences as opposed to posting simply for the sake of it" in the guidelines. I'll shoot off an email to the mods about this later today.
I’d much rather a useful AI-aided post that gives me insight than the almost daily pedantic unrelated gripe “interesting post but I hate the font”. This AI slop is bad but is it worse? In any case “bad” should be a reasonable bar to get over.
Ha! I love this. There's nothing like a proper Bash script to make me realize how terribly gross all of mine are.
The drum I'm currently beating is that local MCP is a ton of fun for techies like us - if you're on this website you can `npx ...` or install whatever you want with a modicum of common sense - but local MCP is going to be a dead end for mass adoption. If we want to build MCP servers that get used by everyday people (or on mobile or other locked down ecosystems) then remote MCP + OAuth is the only realistic way forward. I can't get my dad to open up a terminal window - anything over stdio or touching environment variables and API keys is a nonstarter.
The infrastructure around MCP has a long ways to go before ordinary people can use it. Don’t forget you also have to edit configuration files.
Oh absolutely - but the infrastructure required to support a "click link, get remote MCP URL added to config automatically" flow is _so_ much smaller than the infrastructure required for a "click link, download and install arbitrary software that may or may not depend on having existing tools installed" flow.
Very cool! The docs here are a great overview of how MCP works, and a reminder to me of an old lesson:
We never should have abandoned REST. The whole point was for an interface to be self-describing; we wouldn't need MCP (or Swagger, or OpenAPI, etc) if we just stuck to REST instead of diverting down the JSON RPC route we've been on for 20 years.
Wait, who's abandoned REST?
And in what way is OpenAPI an abandonment of REST? It's an API documentation system that can be leveraged for generating REST server boilerplate code. If anything, it builds up the quality-of-life around REST.
I haven't seen a REST API in production for many years, maybe 15?
That's anecdotal obviously, but almost every, if not every, API I use today is an RPC call returning JSON.
Edit: to be clear, the distinction between what REST was defined as and what we use today often doesn't really matter. We use JSON APIs today, it is what it is. This is a case where it really matters though, LLM companies are now trying to push an entirely new protocol that tries to do roughly what REST did in the first place.
So the things we call "REST" in 2025 are not quite the same as the original specification of REST. One key aspect that has been abandoned is that sent data should be self-describing. That is, it shouldn't require any additional information to be useful. i.e. API documentation for JSON endpoints.
There's a great chapter on this in Hypermedia Systems[1]. Talks about both this and HATEOAS(Hypermedia as the engine of application state).
1. https://hypermedia.systems/components-of-a-hypermedia-system...
By REST you mean HATEOAS?
That's one constraint of REST, yes.
You can't have REST without it
I have to say this is a very readable implementation to see how it all works in practice as well as a good reminder that it's a pretty simple universal tool interface.
>> a good reminder that it's a pretty simple universal tool interface.
That's because it's not really doing anything new. MCP is a land-grab by one company, quickly supported by the rest as they desperately work to abstract and supplant with their own "protocols". Welcome to the era of thin veneers that add little but complexity over what we already had.
Land grab? Not sure I'd call it that, but maybe it is.
I looked at it more as there were N different ways of configuring tools to be called, so they created N+1 (https://xkcd.com/927/) but have had good success at getting alignment on it.
While not knowing the reasons that OpenAI supported it, I would imagine it was along the lines of: "This is not more bad than the others we have seen, may as well use it"
Since that time, thankfully many folks have jumped into looking at it and making it better. I just wish the spec was good and easy to follow (I read through it, and I'm still looking for the real spec)
It works great with Emacs :)
https://github.com/dotemacs/emacs-mcp
I like the fact that it's just Bash
> in pure Bash.
Not really in "pure bash". Also this needs to be labeled as a "toy".
Using an external tool like 'jq' especially written in C for parsing JSON, one can craft a exploitable JSON input to achieve code execution on the MCP server.
What could possibly go wrong? Maybe this CVE-2025-48060 [0] [1]?
[0] https://github.com/jqlang/jq/issues/3327
[1] https://nvd.nist.gov/vuln/detail/CVE-2025-48060
Runtime is called POSIX userspace.
What does zero-overhead mean here?
Raw protocol, really. No marshaling, no conversions, none of the overhead from type management you get with modern Python, none of the turtles-all-the-way-down dependencies of NodeJS equivalents. I like it, although I would probably port it back to “lightweight” Python in about half the size :)
Interesting to see ppl caring about marshalling overhead when working with LLMs
Some of us still prize compute efficiency, especially those who have been using Python for a long time and are contemplating the new kinds of code patterns that have emerged from data science...
Similar to https://github.com/inercia/MCPShell, but the MCPShell can sandbox the execution of the shell code for higher safety.
I don't think they are comparable. MCPShell is a go program to run shell scripts, while the other one allows to define MCP operations as bash functions.
Not quite the same. The bash sdk can't be used to run arbitrary shell commands any more than to run arbitrary python programs.
I just rolled my own Python umcp library based on this, so thanks for the inspiration!
https://github.com/rcarmo/umcp
Gross. I love it.
Did the AI help write this?
I love that “the AI” has become a modern day “the Google”.
"I AI'd it."
here's a node version of an MCP server: https://www.npmjs.com/package/@profullstack/mcp-server
This is neat but "zero runtime" is a misnomer. Bash _is_ runtime, not to mention external tools used in shell scripts like jq.
[flagged]
[flagged]
[flagged]
[flagged]
Based on your recent comment history vibes, I'm pretty sure that you're using a LLM to post comments; and the remainder is self-promotion towards your own articles. Please don't do that here; HN aims to have interesting and not generic conversations: https://news.ycombinator.com/newsguidelines.html
If a comment seems inappropriate for HN (and LLM-generated comments are inappropriate), please flag it and email the mods at hn@ycombinator.com.
fwiw I don't have a problem with LLM posts. But I do agree that this is pretty generic. If you want to use an LLM to post comments: find a better prompt / workflow
LLM-generated comments are not wanted on HN; we want to preserve it as a place for discussion between humans.
Also, comments of the format "here's what an LLM said about this topic:" are best avoided. We don't want to normalise a style of discussing issues in which we generate an LLM output and make that the central conversation topic; we prefer original human thought here.
I guess if we're gonna do monkey's paw/work to rule type of interpretations, HN should just add "we value authentic human experiences as opposed to posting simply for the sake of it" in the guidelines. I'll shoot off an email to the mods about this later today.
I’d much rather a useful AI-aided post that gives me insight than the almost daily pedantic unrelated gripe “interesting post but I hate the font”. This AI slop is bad but is it worse? In any case “bad” should be a reasonable bar to get over.