The Model Context Protocol lets an AI assistant call tools directly. The Inferix MCP server gives Claude, Cursor, or any MCP-compatible client the ability to search the hub and the GPU marketplace itself, instead of asking you to paste results in. ## What it exposes Five tools, all read-only: - `inferix_search_models` — search by free-text query, task, or library - `inferix_get_model` — full detail for one model: card, files, tags, licence - `inferix_search_datasets` — search datasets by query - `inferix_search_gpus` — search rentable machines, filtered by GPU type and maximum hourly price - `inferix_market_metrics` — current marketplace supply, demand, and pricing Read-only is deliberate. The server can look things up; it cannot create, change, or delete anything in your account. An agent that goes wrong with these tools wastes a call — it does not modify your repositories. ## Setting it up It runs locally over stdio. Nothing is hosted and nothing is proxied. Add it to your client's config: ```json { "mcpServers": { "inferix": { "command": "npx", "args": ["-y", "inferix-mcp-server"], "env": { "INFERIX_API_URL": "https://inferix.co/api", "INFERIX_API_TOKEN": "<your-token>" } } } } ``` Restart the client and the five tools appear. The token is optional — public models, datasets, and GPU listings are readable without one. It only adds what is private to your account. ## What it changes in practice **Model selection stops being guesswork.** Ask for a small multilingual embedding model and the assistant searches and reads the cards rather than recalling a name from training data, which is where hallucinated model IDs come from. **Costing happens before the job, not after.** With GPU search and market metrics available, an assistant can check what is actually available at what rate and do the arithmetic with real numbers. **Context stays smaller.** The agent fetches the one model card it needs instead of you pasting several, which matters more than it sounds on long sessions. ## Where it stops It searches and reads. It does not deploy endpoints, start training runs, or rent machines — those are write operations, and putting them behind an agent's judgement is a decision that deserves its own design rather than being bundled into a search tool. For programmatic control of those operations, use the SDKs (`pip install inferix-sdk`, `npm i inferix`), where the call is explicit and yours. The source is about a hundred lines and lives in `mcp-server/` in the Inferix repository. It is worth reading before you connect it to anything — it is short enough to audit in a sitting, and you should know what a tool can do before you grant it to an agent.

Building agents with the Inferix MCP server