# API and MCP for AI agents | What Can I Run AI?

> Free JSON API, OpenAPI spec and MCP server for LLM hardware fit data: which models run on which GPU or Mac.

Source: https://whatcanirunai.vercel.app/developers/

# What Can I Run AI? API for agents and developers

Everything on this site is available as JSON for AI agents, scripts and tools. No key and no sign-up. Use it when a user asks whether an open LLM will run on their GPU or Mac, how much VRAM a model needs, how fast it will run, or what hardware to buy for a model.

## Quick start

curl 'https://whatcanirunai.vercel.app/api/v1/fit?model=qwen3-14b&hardware=rtx-4090&quant=Q4'

## Endpoints

* `GET /api/v1/index.json`: catalog of endpoints and files.
* `GET /api/v1/models.json`: all 656 models with parameters and memory need at Q4, Q8 and FP16 (static, free).
* `GET /api/v1/hardware.json`: all 82 devices with usable memory, bandwidth and list price (static, free).
* `GET /api/v1/fit?model=&hardware=&quant=`: one fit verdict with tokens-per-second range (free).
* `GET /api/v1/what-runs?hardware=`: every model on one device (free).
* `GET /api/v1/recommend?model=&quant=&budget=&min_tps=`: cheapest devices that run a model (free during preview; planned $0.01 per call via x402).
* `GET /api/v1/matrix?quant=`: full model by hardware matrix (free during preview; planned $0.02 per call via x402).

## Machine-readable specs

* [OpenAPI 3.1 spec](https://whatcanirunai.vercel.app/openapi.json)
* [llms.txt](https://whatcanirunai.vercel.app/llms.txt) and [llms-full.txt](https://whatcanirunai.vercel.app/llms-full.txt)
* MCP server (Streamable HTTP): `https://whatcanirunai.vercel.app/mcp`, manifest at </.well-known/mcp>
* Every page is also available as Markdown: send `Accept: text/markdown` or add `index.md` to the URL.

## Authentication

None. All endpoints are public and read-only. Please cache responses; the data changes at most a few times a day.

## Paid calls with x402

Heavier endpoints will use [x402](https://www.x402.org/), the open HTTP 402 payment standard, so agents can pay per call in USDC without an account. When enabled, a call without payment returns HTTP 402 with a `PAYMENT-REQUIRED` header; your x402 client signs and retries with `PAYMENT-SIGNATURE`, and the response carries `PAYMENT-RESPONSE`. You are never charged for an error. Static data and single fit checks stay free.

## Errors

Errors use `application/problem+json` with a `title` and a `hint` link to the list of valid ids.

## Embed widget

Drop a live "what can this GPU run?" checker on a blog or docs page. No key, no signup. The widget is an iframe pointing at </embed/>.

<iframe
  src="https://whatcanirunai.vercel.app/embed/?gpu=rtx-4090"
  title="What Can I Run AI embed"
  loading="lazy"
  style="width:100%;max-width:440px;height:420px;border:0;border-radius:12px;overflow:hidden">
</iframe>

Query params: `gpu` (device id) and optional `n` (GPU count, 1–2). Full interactive reverse lookup: </what-can-i-run/>. JSON: `GET /api/v1/what-runs?hardware=`.

## Accuracy

Numbers are estimates from [our published formula](https://whatcanirunai.vercel.app/method/). Prices are list prices, not live street prices.

---
Agent guide: https://whatcanirunai.vercel.app/llms.txt · JSON API: https://whatcanirunai.vercel.app/developers/