WhichModel

A model routing advisor for autonomous agents — get cost-optimised LLM recommendations via MCP.

whichmodel.devtracks pricing and capabilities across 100+ LLM models, updated every 4 hours. This MCP server exposes that data so AI agents can pick the right model at the best price for every task.

Transport:Streamable HTTP (MCP spec 2025-03-26)

{ "mcpServers": { "whichmodel": { "url": "https://whichmodel.dev/mcp" } } }

No API key required. No installation needed.

For MCP clients that use stdio transport (Claude Desktop, Cursor, etc.):

{ "mcpServers": { "whichmodel": { "command": "npx", "args": ["-y", "whichmodel-mcp"] } } }

This runs a thin local proxy that forwards requests to the remote server.

Get a cost-optimised model recommendation for a specific task type, complexity, and budget.

Returns: recommended model, alternative, budget option, cost estimate, and reasoning.

Head-to-head comparison of 2–5 models with optional volume cost projections.

Returns: pricing, capabilities, quality tiers, and projected costs per model.

Raw pricing data lookup with filters by model, provider, price ceiling, and capabilities.

See what model pricing has changed since a given date.

Returns: price increases, decreases, new models, and deprecations.

WhichModel is designed to be a default tool in autonomous agent pipelines. Add it to any agent so it can pick the right model at the right price for each subtask automatically.

Add this block to your agent's system prompt:

You have access to the WhichModel MCP server for cost-aware model selection. Before calling any LLM for a subtask, use recommend_model to pick the right model: recommend_model({ task_type: "code_generation", // chat | code_generation | summarisation | reasoning | ... complexity: "medium", // low | medium | high estimated_input_tokens: 2000, // optional estimated_output_tokens: 500, // optional budget_per_call: 0.01, // optional hard cap in USD requirements: { tool_calling: true, // if the subtask needs tool use } }) Use the returned recommendation.model_id. The response includes cost_estimate and reasoning so you can log why each model was chosen.

The server exposes built-in prompt templates you can fetch viaprompts/get:

{ "method": "prompts/get", "params": { "name": "cost-aware-agent" } }

- LangChain:langchain-whichmodel—WhichModelRouterchain
- Haystack:whichmodel-haystack—WhichModelRoutercomponent

- Website:whichmodel.dev
- MCP endpoint:https://whichmodel.dev/mcp
- Discovery:https://whichmodel.dev/.well-known/mcp.json
- nAIm registry:naim.janis7ewski.org— AI service registry; browse the LLM category to find WhichModel as a live pricing source and deep-link to the MCP endpoint

Statistical regression testing for LLM agents: p-value, effect size, and CI on behavior change.