If you want people to adopt your tool, you may need to explain any advantage this has over fastmcp.

It's already easy to expose a python function as an MCP server. From the fastmcp docs:

  from fastmcp import FastMCP
  
  mcp = FastMCP("Demo ")
  
  @mcp.tool
  def add(a: int, b: int) -> int:
      """Add two numbers"""
      return a + b
  
  if __name__ == "__main__":
      mcp.run()

Hi thanks for the comment, I’m not trying to replace FastMCP (or anything else), and I’m not really comparing on the “basic MCP server” use case.

PolyMCP, beyond creating MCP servers over HTTP and stdio, WASM (Pyodide) bundle to run tools in the browser/edge with an “MCP-style” tool interface,provides unified agent/orchestration across multiple MCP servers, plus an Inspector UI and production guardrails (budgets, logging, redaction, allowlists, retries).

The goal is to be a single, end-to-end toolkit for developers: tool exposure + debugging + governance + orchestration.

None of this came across in what you wrote in your Show HN.

because I wanted to introduce this part of PolyMCP that can take you from code with only functions to an MCP.

Their docs say the enterprise plan comes with a max of 30 VMs.

I started reading this thinking 'why not just use different port numbers' but I came away convinced that the problem was worth solving and their solution is neat.

> came away convinced that the problem was worth solving

What convinced you? I don't see it. The user is using SSH, if they can't pass a -p option (or type it in a GUI) to their SSH client they won't be able to do much with the shell they're getting either?

I like that you can just use the hostname for web and ssh, without considering that the same IP address isn't exclusively yours.

And, sure, you can add a -p option. But if you have 20 VMs (which is how many come with their basic plan) you'd have to remember all the different port numbers.

(I'm not in the target market for their service.)

hmm. I see the point about using the same hostname… but that's what .ssh/config is for.

You also can't really use the public hostname for this, can you. Unless you do really complex DNS trickery, you can only return one (set of) IP address for a given name. It would thus need to be the same IP address for everyone. Which works only as long as 2 users don't have overlap in the VMs they want to access…

(I guess they can run a solver and try to make it work for as long as possible, including reassigning IPs… but it'll hit a wall at some point?)

Sorry, I don't understand your point about the DNS thing. I don't think multiple owners share the same hostnames. Each owner has a set of 20 hostnames that are unique to their account. And there are 20 IP addresses shared across all owners.

> I don't think multiple owners share the same hostnames.

That's exactly what I mean, this approach wouldn't be able to handle unconstrained sharing of systems among multiple users. If you're, say, a freelancer who has access to a bunch of people's systems… and another freelancer has access to half of those, and then a bunch of others… these combinations create exclusions that can make the whole thing unsolvable if they're large enough.

They're aiming to reduce complexity for the user wherever they can, so their efforts around this make total sense for the platform they've created.

Neat but fragile. It needs a custom proxy and it’s very dependant on specific network setups (eg: doesn’t work in cloud environments).

If you want to do custom voice cloning, record a sample wav file with a sentence or two, and then try this:

  uv tool install --force git+https://github.com/Blaizzy/mlx-audio.git --prerelease=allow
    
  python -m mlx_audio.tts.generate --model mlx-community/Qwen3-TTS-12Hz-0.6B-Base-bf16 --text "Hello, this is a test." --ref_audio path_to_audio.wav --ref_text "Transcript of the reference audio." --play

Brex has been around for a long time, so employees will have been issued stock options with vastly different exercise prices.

Early employees' options will have value, but more recent options are likely underwater.

I strongly suspect they shifted to RSUs at those valuations.

There's double trigger RSUs and so on that allow you to have reasonable tax treatment, due to the theoretical threat of loss if liquidity isn't available. I worked at a company that had this at least while I was there.

I was a regular SDE at brex for a couple years and my various documents about comp say I have RSUs, and carta says so as well.

I've never bothered to understand the details since none of the private companies I've worked for have had the non-cash portion of their comp be worth anything but $0 before.

The usual move here is "double trigger" RSUs that don't vest until a liquidity event, thus no taxes due until said liquidity event.

Right. Plus often the tax is paid out of RSUs given, you just get less in RSUs, some is subtracted to pay tax.

Are those common for regular employees?

highly common

Nope, you as a company owner are highly motivated to shift to RSUs once you hit a certain valuation and number of employees. Everyone does it.

Can you expand on why at a certain valuation and size you would shift?

options can only last 7 years, but brex was founded in 2017.

Who says options can only last 7 years?

ISO options have to expire within 10 years of when they are granted. Sometimes companies make them expire earlier than that, so OP might be thinking of options they were granted. E.g. I once had options that expired 30 days after ending employment even thought the ISO requirement is up to 90 days.

When the options expire do they give new equivalent ones to the employees that hung on? Otherwise what’s the point?

You have to exercise the options or let them expire. You normally have 10 years not 7, but if a company comes up on 10 years after they issued their first options, they might try a tender offer to buy some employee shares. If your 10 year old "start up" shares can't be sold anywhere, then they probably aren't worth exercising. A company that can't provide liquidity to employees for 10 years will probably never do it.

You can exercise the options before that time is up, paying the strike price to convert them to shares

But then you’re potentially stuck with worthless options you can never trade, right? Seems very unfriendly to employees

> Seems very unfriendly to employees

Ding ding ding ding ding!

Most ICs figure this out sooner or later. Unfortunately many only discover it after being screwed hard financially.

Which local model works best with this? (Assuming MacOS with 32GB unified RAM)

gpt-oss 20B works well. You'll want at least 12k context length for agent mode.

Kudos! I love seeing things people have built for their kids.

This reminds me a bit of this site, which has been around for a long time and has a similar motivation: teach physics concepts using simulations:

https://phet.colorado.edu/en/simulations/filter?subjects=phy...

Not the same level - but teaching my kid to use keyboard, and i vibecoded this: https://ai.studio/apps/drive/1MErVnhhK89peNh1RPa3y-ufFExG5Bn...

This is cool!

I'm curious whether you're using this to teach initial familiarity, or as the first step toward touch typing?

For touch typing, I think Typing Club is a good place to start for kids. And then Keybr to develop full fluency. And then Monkeytype to develop speed.

EDIT - when I wrote this comment I had only opened the link on my phone. Now I see it on desktop it's clear that the on-screen keyboard is intended to teach key positions without the user looking down at their keyboard. It's good.

In case anyone else was curious, `ai.studio/apps` apps apparently require a google account login by default. Google ai-mode says allowing public access is an author option, but then (1) API costs are borne by the author, and (2) there are some (unclear) compliance implications.

This is cool

Has anyone successfully run this on a Mac? The installation instructions appear to assume an NVIDIA GPU (CUDA, FlashAttention), and I’m not sure whether it works with PyTorch’s Metal/MPS backend.

FWIW you can run the demo without FlashAttention using --no-flash-attn command-line parameter, I do that since I'm on Windows and haven't gotten FlashAttention2 to work.

It seems to depend on FlashAttention, so the short answer is no. Hopefully someone does the work of porting the inference code over!

Yes, using mlx-audio. See https://news.ycombinator.com/item?id=46726440

Thanks! Simon's example uses the custom voice model (creating a voice from instructions). But that comment led me eventually to this page, which shows how to use mlx-audio for custom voices:

https://huggingface.co/mlx-community/Qwen3-TTS-12Hz-0.6B-Bas...

  uv tool install --force git+https://github.com/Blaizzy/mlx-audio.git --prerelease=allow
    
  python -m mlx_audio.tts.generate --model mlx-community/Qwen3-TTS-12Hz-0.6B-Base-bf16 --text "Hello, this is a test." --ref_audio path_to_audio.wav --ref_text "Transcript of the reference audio." --play

I recommend using modal for renting the metal.

I believe that's something it does automatically now.

In the past it did help a lot, but now it's the default behavior.

I've seen it create plans of 6-8 steps. I have not seen it automatically decide to break down a spec into 30+ tasks.

Ah, this.

Of course, but if your single request to claude code requires 30+ tasks then it's certainly way too big for llm to code from scratch.

This is false.

I've had Claude Code successfully work on a spec that requires 30+ tasks, with my only input being 'great, keep going'.

Excellent!

If LLMs are capable of delivering completely automatically a solution that includes that many tasks then surely in 6 months from now programmers will be out of job.

/s

The CEO's compensation in 2024 was $1.3 million.

That is seven figures, no?