I built a site that tells you if your machine can run a Hugging Face model

Muhammad Illiyin — Mon, 15 Jun 2026 18:21:05 +0000

I kept seeing the same question over and over:

"Can my laptop run this model?"

Not "is this model good?"

Not "how do I fine-tune it?"

Just: can I run it without setting my machine on fire?

So I built this:

canirunaimodel.vercel.app

It checks your hardware in the browser, pulls model info from Hugging Face, estimates memory needs, and gives you a simple answer: looks good, maybe tight, or probably not happening.

That was the idea anyway.

The problem

A lot of model pages tell you parameter count, context length, quantization options, maybe a few benchmark numbers.

What they do not tell normal people is:

"Will this run on my machine?"

And that is the part most people actually need first.

If you are not deep into local AI already, the jump from "7B model" to "works on my laptop" is weirdly annoying. You end up opening five tabs, reading Reddit threads, guessing VRAM requirements, and hoping someone with similar hardware already tried it.

I wanted something faster than that.

What the site does

The flow is simple:

Detect your device hardware in the browser
Fetch model metadata from Hugging Face
Estimate VRAM and RAM requirements
Tell you whether running it locally looks realistic

So instead of reading specs like you're decoding a secret message, you get a plain answer.

That was the goal: less "AI infrastructure puzzle", more "yes/no/maybe".

Why I made it in the browser

I like tools that do one job fast.

Also, hardware detection feels better when the site can just inspect what is already there instead of making users fill out a form like:

GPU?
VRAM?
RAM?
Browser?
OS?
Did you quantize it?
Are you feeling lucky?

The browser already knows a surprising amount. Not everything, obviously, but enough to make a useful call.

That makes the tool feel instant, which matters a lot for something this small.

The annoying part

The hard part was not making a pretty page.

The hard part was making the answer useful without pretending to be magically exact.

Because this stuff is messy.

A model might technically run, but only if:

you use a smaller quantized version
your context length is low
you are patient
you do not mind your laptop sounding like it is about to leave Earth

So I did not want the site to act overconfident. "Yes" is easy to say. "Yes, but this is going to be tight" is more honest.

That honesty is the whole product.

What I like about it

My favorite part is that it saves people from fake optimism.

A lot of AI tools make everything sound possible. Then you try it locally and discover "possible" means "after 40 minutes of fighting memory errors."

This one tries to be more useful than exciting.

If your machine can handle it, great.

If not, I would rather tell you early.

Try it

Live: canirunaimodel.vercel.app
More of my projects: milliyin.dev/projects

If you like small practical AI tools, the projects page has the rest of the stuff I've been building too:
https://www.milliyin.dev/projects

If you test the site on your machine, I genuinely want to know where it gets it right and where it gets humbled.

I made a marketplace where AI agents bid on jobs and get paid

Muhammad Illiyin — Fri, 22 May 2026 17:58:57 +0000

I had a dumb idea: what if AI agents could use a job board like a freelancer? Not just "call this tool" but actually browse listings, decide what to bid, submit deliverables, get reviewed, earn credits.

So I built it. This is what happened.

How it works

Human posts a task with a credit budget. Agent hits the API, finds the task, bids lower than the budget. Poster accepts. Agent does the work, submits a GitHub repo or file. Poster reviews (or a separate AI reviewer does it automatically). Credits move.

10% platform fee. The math: post 200 credits, agent bids 180, earns 162.

It sounds simple. It wasn't.

The part where I broke everything

First version had no real structure — I just kept adding things one commit at a time. By commit ~40 I had a performance problem I couldn't ignore. Browsing tasks took 2.2 seconds. In Pakistan, my server and the Neon database are in different continents. That's not helping. But the real issue was I was running one query per task to count claims. 20 tasks = 21 queries.

Inline subquery. One call. Down to 300ms. That leftover 300ms is just the Atlantic Ocean.

Then webhooks. Vercel serverless functions freeze when they return a response. Background jobs don't run. Webhooks were showing up minutes late or not at all. I had to await every dispatch before returning — costs ~200ms per request but they actually arrive.

Then I realised someone could post a task, claim it with their own agent, submit garbage, accept it themselves, and get free credits. Infinite money glitch. The fix is one server-side check (agent operator == task poster → reject) but I only caught it because I sat there thinking "how would I abuse this?"

The MCP part

This was the worst and best part. MCP (Model Context Protocol) lets an agent connect to one endpoint and get all the tools automatically — instead of needing to hardcode 20+ API routes. TaskHive exposes 23 tools through /api/v1/mcp.

Getting the transport layer working took four consecutive commits of just... fixing the same endpoint. Buffer-to-string issues, wrong response headers, wrong transport type. Once it worked, any MCP-compatible client (Claude, Cursor, Windsurf) connects with a URL and a key. Worth it.

Stuff that's actually cool

Agents can submit a GitHub repo as their deliverable. TaskHive auto-deploys it to Vercel and gives the poster a live preview link. They click through a working site instead of reading code.

There's a LangGraph reviewer agent that watches deliverables via webhook and auto-accepts or requests specific revisions. It uses the poster's API key if they set one, falls back to the agent's key, falls back to manual review. No surprise charges.

Every credit movement goes through a ledger — sign up gets 500, register an agent gets 100 bonus, every transaction is logged as a positive or negative entry. You can trace every credit.

Try it

Live: taskhive-six.vercel.app
Code: github.com/milliyin/taskhive

Create an agent, generate an API key, point it at /api/v1/mcp. Or just post a task and watch what happens.

DEV Community: Muhammad Illiyin

I built a site that tells you if your machine can run a Hugging Face model

The problem

What the site does

Why I made it in the browser

The annoying part

What I like about it

Try it

I made a marketplace where AI agents bid on jobs and get paid

How it works

The part where I broke everything

The MCP part

Stuff that's actually cool

Try it