MCP server / free / no signup

Let your agent pick.
You keep the receipts.

The recommend_model tool answers from the same cited rankings as this site. It is a plain lookup with no language model behind it, so your agent does the thinking: it names the task category and states its limits as numbers, and gets back the same answer every time.

1. Connect

Streamable HTTP at https://modelmatchbook.com/mcp. Stateless, read-only, no key.

Claude Code

claude mcp add --transport http model-matchbook https://modelmatchbook.com/mcp

Cursor, Claude Desktop and other JSON configs

{
  "mcpServers": {
    "model-matchbook": {
      "url": "https://modelmatchbook.com/mcp"
    }
  }
}

2. Ask

An agent building a cheap coding agent with long context would send:

{
  "use_case": "coding-agentic",
  "priority": "price",
  "needs_tool_calling": true,
  "min_context_tokens": 200000,
  "max_input_usd_per_mtok": 3
}

Free text is rejected on purpose. Unknown fields or values come back as an error listing what is allowed, so the agent can correct itself.

Inputs

Only use_case is required.

use_case
One of 47 slugs. Pick the most specific sub-type that fits. The full list with descriptions ships in the tool schema, and each slug matches a page under /best.
priority
Optional re-sort. quality: best measured result. price: cheapest model with real task results. open: open-weight models first.
min_context_tokens
Smallest context window you need, in tokens. Estimate your largest prompt plus expected output yourself.
max_input_usd_per_mtok
Hard cap on input price in USD per million tokens. Convert your budget yourself. Models with unknown pricing are excluded.
max_output_usd_per_mtok
Hard cap on output price in USD per million tokens. Models with unknown pricing are excluded.
needs_image_input
The model must accept image input.
needs_tool_calling
The model must support tool/function calling.
needs_structured_outputs
The model must support structured (JSON schema) outputs.
needs_reasoning
The model must have a reasoning mode.
open_weights_only
Only models with confirmed open weights (self-hostable).
inference_routes
Where you can call the model. A model qualifies if it is listed on any of these. Omit for any route.
exclude_providers
Model providers you must not use (vendor policy).

What comes back

picks
Up to three ranked models with verdict, price, context, weights status and a link to the model page.
top_pick_tip / top_pick_watch_out
A cited strength and caveat for the first pick, with source URLs.
tradeoffs
Strong models that miss your priority, so you can see what you gave up.
near_misses
Ranked models your constraints excluded, each with the exact reason.
data_version
A content hash. Same input and same version always give the same answer.

Rankings follow the methodology. The data refreshes whenever the site rebuilds.