Every tracing tool I tried treated MCP tool calls like generic HTTP — “340ms, status: error” and nothing else. Couldn’t tell me if the tool actually failed, if the client called a tool that doesn’t exist, if the arguments failed validation, or if it was just a 401 (which is a normal part of the MCP auth handshake, not an error).
So I built mcpobs. Drop it into your server (pip install mcpobs, then instrument(mcp)) and it classifies failures into categories that map to an actual fix, tracks progress on long-running tools while they’re still running, and shows the downstream db/http/llm calls in the same trace.
Apache-2.0, self-hostable, docker compose up and you’re running — no account needed: https://github.com/Ruturajs4/mcpobs
I also host it for people who don’t want to run the pipeline themselves — invite-only for now at https://mcpobs.dev, since I’d rather work with a few early users than scale something unstressed.
Would love to hear what’s broken for MCP servers you’re running that I haven’t thought about yet.