Will It Run? MCP server FREE

A remote Model Context Protocol server so AI agents can call the same fit checker as this site. Read-only, no key, no sign-up. Same data and code as the API v1.

Endpoint (Streamable HTTP, stateless): https://willitrunllm.com/api/mcp

Claude Code

claude mcp add --transport http will-it-run https://willitrunllm.com/api/mcp

Cursor

Add to ~/.cursor/mcp.json (or .cursor/mcp.json in a project):

{
  "mcpServers": {
    "will-it-run": {
      "url": "https://willitrunllm.com/api/mcp"
    }
  }
}

Claude Desktop

On paid plans: Settings → Connectors → Add custom connector, URL https://willitrunllm.com/api/mcp. Or use the local bridge in claude_desktop_config.json (needs Node.js):

{
  "mcpServers": {
    "will-it-run": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://willitrunllm.com/api/mcp"
      ]
    }
  }
}

Other clients (LM Studio, OpenClaw, VS Code, …)

Any client that supports remote Streamable HTTP servers: use the URL above. For stdio-only clients, run the bridge: npx -y mcp-remote https://willitrunllm.com/api/mcp.

Tools

Tool	What it does
`search_hardware`	Find GPUs, Apple Silicon Macs and unified-memory machines (Strix Halo, DGX/RTX Spark) the Will It Run? checker knows, with memory, GPU-usable memory and bandwidth. Use the returned id as `hw` in check_fit / what_fits.
`search_models`	Find open-weight LLMs in the Will It Run? database (params, MoE, release date, max context). Use the returned slug as `model` in check_fit.
`check_fit`	Will this model run on this hardware? Returns verdict (fits / tight / offload / no), memory needed vs GPU-usable memory (estimates), placement for RAM offload (MoE experts-to-RAM like llama.cpp --cpu-moe, or layer split), measured or estimated generation speed, and a checker deep link. Give `model` (slug) OR `hf` + `file` (any Hugging Face GGUF repo/file, sized from the real file). Omit `quant` to get the best quant that fits.
`what_fits`	List the models that fit on this hardware (best quant each), fully on GPU first, then with RAM offload, biggest first. Includes speed where known and checker deep links.
`get_benchmarks`	Measured community llama.cpp benchmarks (tg128 = generation tok/s, pp512 = prompt tok/s), each with its source link. Filter by hardware id, model slug and/or quant.

Example

Ask your agent: “Can my RTX 5060 8GB with 32 GB RAM run Qwen3.6 35B A3B at 96k context with q8_0 KV cache?” It calls check_fit with {"hw":"rtx-5060","model":"qwen3-6-35b-a3b","ram":32,"ctx":98304,"kv":"q8_0"} and gets the verdict, memory needed, offload placement, speed (measured or estimate) and a link back to the checker.

Limits & notes

Free, best-effort ~120 tool calls/minute per IP (shared with the API).
Memory and speed values marked ESTIMATE are estimates; “measured” speeds are community llama.cpp benchmarks with source links.
Hugging Face GGUF checks (hf + file) look up the real file size through our cached proxy.