A coding agent can pull the data, train in your workspace and submit the predictions in a single session; a chat-only client fetches through the signed-URL tools and submits inline from a dataframe. The hosted server stores nothing — your key authenticates each request exactly as the Python client does.
The CrowdCent MCP server, which now ships inside the crowdcent-challenge package rather than from this repository. It downloads challenge data, submits predictions, reads back how they scored, and backtests portfolio constructions on the community meta-model.
- Signed download URLs from get_training_dataset_url, get_inference_data_url and get_meta_model_url, usable by any agent with a terminal or a fetch tool
- Prediction submission, plus your recent submissions and how each one scored
- Simulation over the meta-model with the same engine and tier gates as the site's Simulation tab — sweep configurations, blend sleeves, compare in-sample against out-of-sample
- The local server additionally carries the file-download and file-submission tools
- switch_challenge is gone: challenge-scoped tools now take an optional challenge_slug, defaulting to hyperliquid-ranking
Two paths, both replacing a local clone of this repository. Hosted: add https://mcp.crowdcent.com/mcp to any client that supports remote MCP over HTTP, with your API key as a bearer token — it is a stateless pass-through that stores nothing. Local: one uvx line, uvx --from crowdcent-challenge[mcp] crowdcent-mcp, with CROWDCENT_API_KEY in the env block; no clone and no absolute paths. Keys come from your crowdcent.com profile settings. If a config still points at a local server.py, replacing that one entry is the whole migration.
One command plus a key — claude mcp add --transport http crowdcent https://mcp.crowdcent.com/mcp --header "Authorization: Bearer YOUR_API_KEY", then supply credentials