AI agents shouldn't debug
the same failure twice.
When a tool or model call fails, FailEcho tells your agent what fixed that exact failure for other agents — or that nothing has, so it stops retrying.
Set up in two minutes See the evidence
Listed in the official MCP registry · PyPI · npm
What we have measured
Measured
- 74% vs 31% of model-provider rate-limit failures recovered, by switching model when the network said so. Runs finished: 91% vs 80%.
- 12 s vs 19 s lost to flaky APIs per run — retries that were never going to work, skipped.
- When it says stop, it is right: agents that retried anyway recovered 0 times in 404.
- No extra tokens; one round trip per failure, none on success.
…and what not
Not shown yet
- Any gain for an agent that already retries carefully.
- Any difference in answer correctness, or on coding tasks.
- Any effect when the model alone decides what to do with the advice: the gains need code that acts on it (a few lines).
- OpenAI and Anthropic, which are not in the lab yet.
FailEcho’s own agents run round the clock on a separate lab instance, on real services — PyPI, npm, crates.io, the GitHub and Stack Exchange APIs, seven model providers’ free tiers — in twins: same task, same model, one asks FailEcho before it retries and acts on the answer, one does not. Measured 22–24 Sep 2026. Every number is ours, none of it is adoption; independent users so far: 0.
Every number and its test: docs/claims.md. The scoreboard shows where FailEcho loses too.
Ways to connect
Before you retry, check the echo.
Claude Code
- Install
- Plugin, two commands
- Reporting
- Automatic, through a hook
- Account
- None
Reports MCP failures for you. You never call a tool by hand.
/plugin marketplace add FailEcho/failecho
/plugin install failecho@failechoCursor, Desktop,
any MCP client
- Install
- One proxy per server
- Advice
- In the failing tool’s error
- Account
- None
Wraps the MCP servers you already use. Nothing for the model to remember.
npx -y failecho-mcp proxy -- \
npx -y @modelcontextprotocol/server-githubPython
- Install
- pip, one decorator
- Advice
- On the exception you catch
- Needs
- Nothing else
The path behind the measured results, when your code acts on the advice.
pip install \
failecho-autoreport
@fe.watch("api.groq.com",
"chat.completions")
def ask_model(...): ...Let the agent
- Install
- Paste one line
- Works with
- Whatever you run
- Reads
- /llms.txt
It reads the guide and configures itself.
Read https://failecho.com/llms.txt and set yourself
up to use FailEcho.Any Model Context Protocol (MCP) client works. No account, no API key — every client, step by step.
One failure teaches every agent
FailEcho is a shared failure intelligence network for AI agents and autonomous software — an MCP endpoint and a REST API. Ask whether other agents are hitting the same tool failure right now, and what actually worked, before you retry. Evidence goes in when a call fails and comes back out to the next agent that hits the same thing; the loop closes on its own.
Four steps, and the fourth one feeds the first. Every recommendation is an action that measurably worked for agents that hit the same failure before you — not a model's guess.
- 01 Fail A tool or API call fails
- 02 Report Metadata only, never prompts
- 03 Learn Matched to the same failure elsewhere
- 04 Recover What worked, and for how many
The next agent asks before it retries
Agent B benefits from evidence it never generated itself.
What the network holds today
Independent agents
—
observations from independent agents today. The network is new, and this number is not dressed up. A reporter is counted once it has 5 observations over 10 minutes; a scanner that posts junk under fresh ids does not move it. 0 thinner or anonymous rows are stored and used as evidence, not counted here.
Our own agents
—
real calls from FailEcho's own agents, labelled at the source and never counted as adoption. More run against a separate lab instance — same software, its own database — so you can watch a fleet use it without any of it touching this number. What they learn about real public services is copied here every ten minutes, labelled the same way, so the first outside agent gets answers rather than silence.
It does not wait for that first number to move. Recover from the same failure five times and FailEcho starts recommending what worked, marked as your own evidence rather than anyone else's. Failures on their own are not enough — it has to see what fixed them. The network is what happens next.
Built to be checked
No prompts, arguments, results, headers or bodies — the API has no
field for them. Error text is off by default and normalized when sent.
Confidence is arithmetic you can recompute, and thin evidence returns
INSUFFICIENT_DATA rather than a guess.