The design decision worth knowing is that the endpoints are not tools: the agent searches the catalog, reads one schema and invokes it, so the tool list stays small however much coverage sits behind it — at the cost of two round trips before the first answer. What comes back in exchange is provenance, with the provider, the retrieval time, the cache age and the fallback state attached to every result.
A hosted MCP server that normalises ten upstream crypto data providers into a single catalog. It does not put every endpoint in front of the model: the agent searches the catalog or lists a category, reads the exact input schema for the endpoint it found, then invokes it.
- search_tools searches the whole catalog and its canonical task toolsets by keyword, provider or category, paginated with a cursor
- get_api_endpoint_schema returns the exact input schema for any endpoint by name, so the call is built from real parameters instead of guesses
- invoke_api_endpoint executes any endpoint in the catalog with the arguments you assembled
- Ten category listers for browsing rather than searching: Market Data and Price, On-Chain DEX & Pool Analytics, Portfolio & Wallet, Token & Contract data, DeFi Protocol Analytics, NFT Analytics, Security & Risk Analysis, Network & Infrastructure, Search & Discovery, Prediction Markets
- Behind those categories sit Alchemy, CoinGecko, DeFiLlama, Moralis, Codex, GoPlus, Helius, Tenderly, CCXT and Hyperliquid
- Every result carries a receipt: which provider answered, Hive's retrieval and first-observation time, cache age, and whether the data came from the live or the fallback tier — fallback and degraded data are labelled rather than blended in
- Time-series endpoints accept at or block_number, so a historical question is pinned to a point in time instead of quietly answering with the latest value
- response_format switches any of these between a concise and a detailed result
A remote Streamable HTTP endpoint at https://mcp.hiveintelligence.xyz/mcp. Interactive clients discover it from the URL and authorise in the browser; headless ones send Authorization: Bearer with a Hive API key. You can also connect with no credential at all — the anonymous lane answers 25 material calls per IP per day, and discovery and schema calls are always free. To run it locally instead, launch hive-intelligence over stdio with npx and supply your own upstream provider keys (COINGECKO_PRO_API_KEY, ALCHEMY_API_KEY, HELIUS_API_KEY, MORALIS_API_KEY, CODEX_API_KEY, GOPLUS_APP_KEY and GOPLUS_APP_SECRET, TENDERLY_API_KEY); providers you leave unkeyed stay discoverable and report a missing_key status.
One command — claude mcp add --transport http --scope user hive https://mcp.hiveintelligence.xyz/mcp