Repository navigation
First public release of FailEcho, a shared failure intelligence network for AI agents.
An agent reports a tool failure; the server normalizes the message into a fingerprint; agents report which recovery they tried and whether it worked. The next agent to hit that fingerprint gets the evidence instead of guessing.
Connect
Remote MCP server (streamable HTTP, no auth, no key):
https://failecho.com/mcp
Four tools: check_tool_failure, report_tool_failure, report_tool_success, report_recovery_outcome. A plain REST API is available at the same origin, documented at https://failecho.com/docs.
What's in it
- Deterministic error normalization — regex only, no model — so equivalent errors collapse to one fingerprint.
- Confidence as a Wilson score lower bound over observed attempts, capped at 0.99 and discounted below three distinct reporters. Recomputable from the counts returned.
- Successes recorded alongside failures, so failure rates mean something.
- Per-reporter evidence caps and write rate limiting.
- Raw observations rolled into hourly aggregates and deleted after 48 hours.
Privacy
No prompts, API keys, Authorization headers, cookies, tool arguments, tool results, or request/response bodies are collected. Raw error text is normalized and then discarded; only the fingerprint is stored. Reporter IDs are optional and salted-hashed.
Running it yourself
Python 3.11+, FastAPI and SQLite in WAL mode. One process, about 100 MB of RAM. See the README for setup.
The public network is new and currently empty.