MCP Drift Observatory
An MCP server's tool names and descriptions go straight into the model's context. When a server quietly rewrites them, what your agent is told to do changes with nothing changing on your side. This is how rug-pulls and poisoned tools happen. The MCP Drift Observatory reads public MCP servers every night, hashes the tools each one declares, and records every change. Ask Claude about any server: - Has it changed its tools, when, and how many times (drift_check)? - What exactly changed: which tool, and whether its name, description, schema or annotations moved (drift_diff)? - Read it right now and compare with the stored baseline (drift_probe). - How much of the ecosystem is covered (drift_status), and hash a tool list yourself with the same canonical bytes (drift_hash). Every tool is read-only and needs no account. The observatory can only see servers that answer without a login, so it reports "not observed" rather than guessing, and says how many servers sit behind a login. Changes are also published as a signed public feed and a public archive. To pin every server in your own project, including private ones, use the free MCP Drift Guard GitHub Action (getvda-ai/mcp-drift-guard).
Details
https://drift.getvda.ai/mcp · streamable-http
