Failecho

Failure intelligence for AI agents and autonomous software. Before you retry, check the echo.

Website · Connect an agent · API reference · Live network · llms.txt

FailEcho is a cross-agent failure intelligence network. When a tool or model call fails, it tells your agent what fixed that exact failure for other agents -- or that nothing has, so it stops retrying. Agents share the shape of their failures and what fixed them (metadata only, never prompts or data); the next agent to hit the same failure gets the answer. Open source, no account.

What we have measured, and what we have not

Our own agents run in twins in a lab: same task, same model, one asks FailEcho before it retries and acts on the answer, one does not. Measured 22-24 September 2026. Independent users so far: 0. Every number, its sample and its significance test: docs/claims.md; live: the lab scoreboard.

Contents

Start here

Connect something

How it works

Run and operate it

Agent A fails.
FailEcho learns.

Agent B encounters the same failure.
It sees what actually worked for other agents.

Agent B benefits from evidence it never generated itself.

Connect in one minute

Easiest: let the agent do it. Paste this at whatever you are running -- Claude Code, Claude Desktop, Cursor, Codex, your own harness:

Read https://failecho.com/llms.txt and set yourself up to use FailEcho.

It reads the machine-readable guide and configures itself. No account, no API key, nothing to sign up for. Everything below is the same thing done by hand.

Put it in the tool path, not the tool list. A tool the model has to choose to call is one it mostly does not call: in our lab, agents given FailEcho's MCP tools used them about once every five runs. The integrations that work hand the model the answer where it is already looking, so pick by client:

All four deliver the advice, and all four are tested end to end. The gains the lab has measured come from agents that act on it -- switch model when told switch_model, stop when told skip (three lines with the wrapper); none of the four integrations has a lab comparison of its own with a result to quote yet.

The bare MCP endpoint -- the smallest option and the least effective

https://failecho.com/mcp

Most MCP clients take this config block:

{
  "mcpServers": {
    "failecho": { "type": "http", "url": "https://failecho.com/mcp" }
  }
}

In Claude Code that file is .mcp.json; Cursor uses .cursor/mcp.json and drops the type; VS Code uses .vscode/mcp.json and calls the top-level key servers. The endpoint never changes. The setup page has the table.

On Claude Code the CLI writes that same file for you:

claude mcp add --transport http --scope project failecho https://failecho.com/mcp

Python, if you want failures and successes reported automatically:

from failecho import FailEcho

echo = FailEcho("https://failecho.com", reporter_id="my-agent-1")

outcome = await echo.observe_tool_call(
    service="github-mcp",
    operation="create_issue",
    call=lambda: github.create_issue(**args),
)

if outcome.failed and outcome.decision.actionable:
    do(outcome.decision.recommendation)   # your code decides, never FailEcho

No account. No API key. Free during the public MVP. Full integration guide: Connect an agent.

What it does

See whether other AI agents are hitting the same tool failure right now — and which recovery actions actually worked. FailEcho exposes a Model Context Protocol (MCP) endpoint that agents can query after a tool failure, plus a REST API.

FailEcho normalizes error text deterministically (no model) into a fingerprint, accumulates recovery outcomes against it, and returns a recommendation only when independent reporters agree. Thin evidence returns INSUFFICIENT_DATA rather than a guess. Confidence is a Wilson score lower bound you can recompute from the counts returned beside it.

It stores failure metadata only. There is no field for prompts, tool arguments, tool results, request or response bodies, headers or cookies, so none of it can be stored. One field is free text, the error message: optional, off by default in the hook and the wrapper, and when sent it is normalized -- identifiers replaced, credential-shaped strings redacted -- and the raw text discarded. That normalization is a second line of defence, not a guarantee; the honest claim is metadata only, error text off by default, normalized when on.

Live: https://failecho.com · /docs · /openapi.json · /llms.txt

This is not an observability platform, an error database, an uptime monitor or an LLM debugger. The unit of the system is:

service + operation + version + schema_hash + failure fingerprint
                     + observed recovery outcomes

Vocabulary

The brand vocabulary is for humans. Wire formats are deliberately unbranded: endpoint paths, MCP tool names and field names (fingerprint, recommendation, recovery_actions) stay exactly as they are, because machine clarity outranks naming purity.

See the network effect locally

Two terminals, about a minute.

# 1. the network
uv run uvicorn app.main:app --reload
#    or: .venv/bin/python -m uvicorn app.main:app --reload

# 2. six independent agents hitting the same broken tool
uv run python examples/live_agent/run_demo.py
#    or: .venv/bin/python examples/live_agent/run_demo.py

The demo starts a small local tool server, then runs six logically independent agents against it. Every network call goes over MCP, from an external process, using the official MCP SDK.

Agent A calls a tool. It fails: the provider renamed a field.
        |
        v
Agent A reports the failure          -> the network records it
Agent A has no evidence to go on, so it retries (fails),
        refreshes the tool schema (works), and reports both outcomes
        |
        v
Agents C, D, E, F hit the same failure with different repository ids
        -> normalization collapses all of them onto ONE fingerprint
        -> the network accumulates evidence from 5 independent reporters
        |
        v
Agent B hits the same failure with yet another id, and asks first
        -> the network recognises the fingerprint
        -> "refresh_schema: 5/5 successes, 5 reporters, confidence 0.57"
        -> "retry: 0/5. Do not bother."
        |
        v
Agent B skips the retry the others wasted a call on, refreshes, succeeds,
and reports its outcome -- which makes the next agent's answer better.

Agent B never met Agent A. It only met the network. That is the entire product.

Real output from the sixth agent, which had reported nothing before it asked:

Calling tool...
x tool failed

  422 validation_error
  Repository 987654 rejected field body: field "body" is no longer accepted, use "content"

Checking shared failure intelligence...

  Fingerprint:            6ed9ef705ff4037af2c977306b8b9f92
  Known failure:          YES
  Observed failures:      11
  Independent reporters:  6
  Service status:         MAJOR

  Recovery actions others reported:
    refresh_schema        5/5 (100.0%) confidence 0.57 reporters 5
    retry                 0/5 (0.0%) confidence 0.00 reporters 5

Best observed recovery:
  refresh_schema
  Skipping retry: other agents already proved it does not work here.

Applying recovery: refresh_schema
  Refreshed tool schema -> v3.0.0, field 'content'
  Retrying tool call...
  + tool call succeeded

Reporting recovery outcome...
+ accepted   (refresh_schema -> success)

Watch it land on the homepage at http://localhost:8000 while the demo runs. Demo agents label themselves with X-Reporter-Kind: demo, so their traffic is real evidence but is never counted as adoption — see Demo data.

Details, including how to run the tool server separately, are in examples/live_agent.

Connect an agent

Two ways in, and the difference matters.

MCP lets an agent explicitly ask and report — the model decides when to call check_tool_failure, so you get intelligence exactly where the agent reasons about a failure, and nothing else.

SDK instrumentation reports success and failure telemetry automatically for every tool call, without the model deciding anything. That is what produces denominators, and without denominators every failure rate in the network is meaningless.

Most deployments want both.

1. MCP

claude mcp add --transport http failecho https://failecho.com/mcp

{
  "mcpServers": {
    "failecho": {
      "type": "http",
      "url": "https://failecho.com/mcp"
    }
  }
}

2. Python

Copy client/ into your project (not published to PyPI yet), then:

from failecho import FailEcho

echo = FailEcho(
    endpoint="https://failecho.com",
    reporter_id="my-agent-1",       # optional, hashed server-side
)

outcome = await echo.observe_tool_call(
    service="github-mcp",
    operation="create_issue",
    version="2.8.1",
    schema_hash="a817ce",
    call=lambda: github.create_issue(**args),
)

if outcome.failed and outcome.decision.actionable:
    # YOUR code decides. FailEcho never acts on your behalf.
    if outcome.decision.confidence > 0.8:
        refresh_schema()
        await echo.report_recovery(
            fingerprint=outcome.decision.fingerprint,
            action="refresh_schema",
            successful=True,
        )

observe_tool_call reports the success or the failure, queries FailEcho when the call failed, and hands you a FailureDecision. It never retries, never refreshes and never falls back — executing a recovery can double-post or double-charge, so that decision stays yours.

It cannot break your agent. Every call is fail-soft: a timeout or an unreachable host is swallowed and your tool result is returned anyway. Set FAILECHO_DISABLED=1 and the whole client becomes a no-op.

3. Framework instrumentation

Reference integration, Pydantic AI:

from failecho import FailEcho
from failecho.integrations.pydantic_ai import instrument_toolset

echo = FailEcho("https://failecho.com", reporter_id="my-agent-1")
agent = Agent("openai:gpt-4o", toolsets=[instrument_toolset(my_toolset, echo)])

Every tool call now reports its outcome. The wrapper is behaviourally invisible: same results, same exceptions, same control flow. Tool arguments are never read and never sent.

Other frameworks (LangChain, LlamaIndex, CrewAI, OpenAI Agents SDK, Claude Code hooks) are not built yet. They should implement failecho.adapters.ToolTelemetrySink — four events, one direction — rather than touch FailEcho's core. See client/failecho/adapters.py.

4. REST

curl -X POST https://failecho.com/v1/query \
  -H "Content-Type: application/json" \
  -H "X-Reporter-ID: my-agent-1" \
  -d '{
    "service": "github-mcp",
    "operation": "create_issue",
    "error_type": "validation_error",
    "error_code": "422",
    "error_message": "Repository 555812 was not found"
  }'

5. Claude Code plugin (automatic)

Connecting the MCP server leaves it to the model to call FailEcho when a tool fails, and models forget. The plugin removes the decision:

/plugin marketplace add FailEcho/failecho
/plugin install failecho@failecho

That installs the MCP server and a hook Claude Code runs after every MCP tool call, so every failure is reported, successes give the failure rates their denominator, and a second attempt is recorded as a recovery (retry with the same arguments, adjust_arguments with new ones). When the network already knows a failure, the hook hands Claude a short note -- how often others hit it and which recovery worked -- before it retries.

Without the plugin, the hook is one file with no dependencies beyond Python 3:

mkdir -p ~/.claude/hooks
curl -fsSL https://raw.githubusercontent.com/FailEcho/failecho/main/plugin/hooks/failecho_hook.py \
  -o ~/.claude/hooks/failecho_hook.py

Then add to ~/.claude/settings.json:

{
  "hooks": {
    "PostToolUseFailure": [{"matcher": "mcp__.*", "hooks": [
      {"type": "command", "command": "python3 ~/.claude/hooks/failecho_hook.py", "timeout": 10}]}],
    "PostToolUse": [{"matcher": "mcp__.*", "hooks": [
      {"type": "command", "command": "python3 ~/.claude/hooks/failecho_hook.py", "timeout": 10}]}]
  }
}

What leaves your machine: the server's public name and the tool name, a coarse error class and code (rate_limit / 429), and the call's latency. Never tool arguments, tool results, prompts, file paths or session ids, and the error text only if you set FAILECHO_HOOK_SEND_ERRORS=1. A server is named by its public package (npx @scope/server, uvx server) or its public host; local scripts and private hosts are skipped entirely. Name one yourself with FAILECHO_HOOK_SERVICE_NAMES='{"alias": "public-name"}'. If FailEcho is unreachable, the hook gives up after one short timeout and Claude carries on.

About reporter IDs

Optional, and never required. A stable one is salted and hashed on arrival — the raw value is never stored — and it improves three things: independent reporter counting, poisoning resistance, and FailEcho's ability to tell you that a recommendation came from somebody other than you. Anonymous reporting stays fully supported.

Your own agents (first-party)

While the network bootstraps, the operator's own agents report real failures too. That data is real field evidence, but it is not independent and it is not adoption, so it carries its own label everywhere it appears:

first_party is a claim about who is reporting, so it has to be proven: send X-FailEcho-Operator: <FIN_FIRST_PARTY_TOKEN>. A wrong or missing token is stored as demo, which keeps it out of adoption and never shows it to anyone as operator evidence. Every query answer lists evidence_sources, so an agent can tell an answer backed only by first_party from one that independent agents back.

Generate the token once, on the server:

echo "FIN_FIRST_PARTY_TOKEN=$(openssl rand -hex 32)" >> /etc/failecho.env

Then give it to your own agents, and nobody else:

# Claude Code
claude mcp add --transport http failecho https://failecho.com/mcp \
  --header "X-FailEcho-Operator: <token>"

# stdio relay
FAILECHO_OPERATOR_TOKEN=<token> failecho-mcp

The Python client takes operator_token="<token>", or reads FAILECHO_OPERATOR_TOKEN.

Naming what failed

The name is part of the fingerprint, so evidence is only shared when agents name the same thing the same way. Use the MCP server's own name (its serverInfo.name) or the HTTP API's host as service, and the tool name exactly as the server defines it as operation: create_issue, not mcp__github__create_issue.

Concept

Agent A fails
    |
    v
reports anonymously  ---------> network learns
                                    |
Agent B hits the same problem       |
    |                               |
    v                               v
queries the network  <---------  what happened to others
    |
    v
skips the useless retry, uses the recovery that works

Run locally

Python 3.11+.

# with uv
uv venv
uv pip install -r requirements.txt
uv run uvicorn app.main:app --reload

# or plain venv + pip
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
.venv/bin/python -m uvicorn app.main:app --reload

Seed synthetic demo data so the homepage has something to show:

.venv/bin/python scripts/seed_demo.py            # add demo data
.venv/bin/python scripts/seed_demo.py --reset    # replace existing demo data
.venv/bin/python scripts/seed_demo.py --purge    # remove demo data

Then:

• homepage — http://localhost:8000
• MCP endpoint — http://localhost:8000/mcp (Streamable HTTP)
• agent-readable overview — http://localhost:8000/llms.txt
• API docs — http://localhost:8000/docs
• machine-readable schema — http://localhost:8000/openapi.json
• health — http://localhost:8000/health

Run the tests:

.venv/bin/python -m pytest

Fold expired raw observations into hourly aggregates (safe to run any time):

.venv/bin/python scripts/prune.py --dry-run
.venv/bin/python scripts/prune.py

End-to-end examples (server must be running):

.venv/bin/python client/example_agent.py            # REST, single agent
.venv/bin/python examples/live_agent/run_demo.py    # MCP, six agents, network effect

The demo runs its tool server in a background thread. To run it separately (two terminals) instead:

.venv/bin/python examples/live_agent/tool_server.py
.venv/bin/python examples/live_agent/run_demo.py --no-tool-server

MCP

The MCP server runs inside the same FastAPI process — no second service to deploy or supervise — and speaks Streamable HTTP at /mcp. It is stateless with JSON responses: no per-session memory, no long-lived streams, which is what keeps it viable on a small VPS.

Connect

Claude Code:

claude mcp add --transport http failecho https://failecho.com/mcp
# local:
claude mcp add --transport http failecho http://localhost:8000/mcp

Generic MCP client config (mcpServers style):

{
  "mcpServers": {
    "failecho": {
      "type": "http",
      "url": "https://failecho.com/mcp"
    }
  }
}

Raw JSON-RPC, if you want to see it work:

curl -s localhost:8000/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Local stdio server

Some hosts can only start a local process and talk to it over stdin/stdout. failecho-mcp is for them. It is a relay, not a second FailEcho: it has no database and stores nothing. Every tools/list and tools/call is forwarded to the shared network, so it serves the same four tools, with the same descriptions and the same evidence, as the URL above.

uvx failecho-mcp

{
  "mcpServers": {
    "failecho": {
      "command": "uvx",
      "args": ["failecho-mcp"]
    }
  }
}

failecho-mcp is published separately from this repository and depends on mcp alone -- 29 packages, about 48 MB, roughly two seconds on a cold cache. The server package (failecho-server, this repository) pulls FastAPI, SQLAlchemy and uvicorn because it is the server; the relay imports none of them.

The same relay exists for Node, in npm-relay/:

{
  "mcpServers": {
    "failecho": {
      "command": "npx",
      "args": ["-y", "failecho-mcp"]
    }
  }
}

Zero dependencies, 6 KB, about a second from a cold npx cache. Same four tools, same evidence, stores nothing. Use whichever runtime you already have.

To run the relay from a checkout while working on it:

uv run failecho-mcp

If the network is unreachable, a tool call returns an error result that says so and records nothing, and the agent falls back to its own retry policy instead of hanging.

Prefer the URL when your client supports it: one hop fewer, nothing to install.

Tools

All four call the same functions as the REST endpoints (app/core/service.py), so an MCP client and a curl user can never disagree about what a failure means — there is one normalizer, one fingerprint function, one intelligence layer.

Example check_tool_failure result:

{
  "known": true,
  "fingerprint": "01ae47053fbb3eabf8f3e480cba45ba8",
  "status": "MAJOR",
  "observations": { "total": 418, "last_5m": 81, "last_1h": 201, "unique_reporters": 47 },
  "failure_rate": { "last_5m": 0.73, "last_1h": 0.31 },
  "recovery_actions": [
    { "action": "refresh_schema", "attempts": 124, "successes": 117,
      "success_rate": 0.9435, "effective_attempts": 124, "unique_reporters": 45,
      "confidence": 0.8881 }
  ],
  "recommendation": { "action": "refresh_schema", "confidence": 0.8881 },
  "demo_data_included": false
}

demo_data_included tells an agent when synthetic demo rows are part of the numbers. Disable MCP entirely with FIN_MCP_ENABLED=0.

REST API

Three calls. No account, no API key, no payment.

Report a failure

curl -s localhost:8000/v1/observe \
  -H 'content-type: application/json' \
  -H 'X-Reporter-ID: my-agent-1' \
  -d '{
    "service": "github-mcp",
    "operation": "create_issue",
    "version": "2.8.1",
    "schema_hash": "a817ce",
    "outcome": "failure",
    "error_type": "validation_error",
    "error_code": "422",
    "error_message": "Repository 918272 was not found",
    "latency_ms": 421
  }'

{
  "accepted": true,
  "fingerprint": "01ae47053fbb3eabf8f3e480cba45ba8",
  "known": true,
  "observations": 143,
  "normalized_error": "Repository <N> was not found"
}

The message is normalized before anything is stored: Repository 918272 was not found → Repository <N> was not found. The fingerprint is sha256(service | operation | version | schema_hash | error_type | error_code | normalized_error), truncated to 32 hex chars.

Report a success

Failure rates need a denominator, so send successes too:

curl -s localhost:8000/v1/observe \
  -H 'content-type: application/json' \
  -d '{
    "service": "github-mcp", "operation": "create_issue",
    "version": "2.8.1", "schema_hash": "a817ce",
    "outcome": "success", "latency_ms": 318
  }'

Query the network

curl -s localhost:8000/v1/query \
  -H 'content-type: application/json' \
  -d '{
    "service": "github-mcp",
    "operation": "create_issue",
    "version": "2.8.1",
    "schema_hash": "a817ce",
    "error_type": "validation_error",
    "error_code": "422",
    "error_message": "Repository 555812 was not found"
  }'

{
  "known": true,
  "fingerprint": "01ae47053fbb3eabf8f3e480cba45ba8",
  "status": "MAJOR",
  "looks_new": false,
  "observations": { "total": 418, "last_5m": 81, "last_1h": 201, "unique_reporters": 47 },
  "failure_rate": { "last_5m": 0.73, "last_1h": 0.31 },
  "recovery_actions": [
    { "action": "refresh_schema", "attempts": 124, "successes": 117,
      "success_rate": 0.9435, "confidence": 0.8881 },
    { "action": "retry", "attempts": 91, "successes": 17,
      "success_rate": 0.1868, "confidence": 0.12 }
  ],
  "recommendation": {
    "action": "refresh_schema", "confidence": 0.8881,
    "based_on_attempts": 124, "based_on_successes": 117
  }
}

When the network has nothing useful:

{ "known": false, "status": "INSUFFICIENT_DATA", "recommendation": null }

/v1/query is read-only. It stores nothing.

Report a recovery outcome

curl -s localhost:8000/v1/outcome \
  -H 'content-type: application/json' \
  -d '{
    "fingerprint": "01ae47053fbb3eabf8f3e480cba45ba8",
    "action": "refresh_schema",
    "successful": true
  }'

{ "accepted": true }

Actions are free-form strings in V1. Common ones: retry, wait, refresh_schema, remove_optional_field, reconnect, use_fallback, reauthenticate, abort.

Status

curl -s localhost:8000/v1/services              # per service/operation health, worst first
curl -s localhost:8000/v1/stats                 # counters; real and synthetic kept separate
curl -s localhost:8000/v1/recovery-intelligence # best evidenced recovery actions
curl -s localhost:8000/health                   # {"status":"ok"}
curl -s localhost:8000/llms.txt                 # agent-readable description of the service

Python client

Zero dependencies — standard library only. Copy client/failure_network.py and client/failecho.py into your agent (the package is not published yet).

failecho is the preferred import name and simply re-exports failure_network, which keeps working unchanged — the rename is additive, so no existing code breaks.

Error text is opt-in. The wrapper's default classifier reports the exception class and status code and no message; set FAILECHO_SEND_ERRORS=1 to send the text as well. error_message passed explicitly, as in the example below, is always sent -- that is your call, not a default.

from failecho import Client   # or: from failure_network import Client

client = Client("http://localhost:8000", reporter_id="my-agent-1")

client.observe_failure(
    service="github-mcp",
    operation="create_issue",
    version="2.8.1",
    schema_hash="abc",
    error_type="validation_error",
    error_code="422",
    error_message="Repository 91827 not found",
)

intel = client.query(
    service="github-mcp",
    operation="create_issue",
    version="2.8.1",
    schema_hash="abc",
    error_type="validation_error",
    error_code="422",
    error_message="Repository 12345 not found",
)

if intel["recommendation"]:
    action = intel["recommendation"]["action"]     # e.g. "refresh_schema"
    client.report_recovery(
        fingerprint=intel["fingerprint"], action=action, successful=True
    )

client.observe_success(service="github-mcp", operation="create_issue", latency_ms=318)

Every call is fail-soft: a timeout or an unreachable server returns None (or a neutral INSUFFICIENT_DATA dict from query) instead of raising. Telemetry must never break the agent it observes.

What connecting asks of you

Nothing. The hosted MCP endpoint (https://failecho.com/mcp), the REST API, the plugins, the proxy and failecho-autoreport need no key, token or account, and read no credential. Every optional setting is listed under Configuration; the only secret any of them takes is your own team token, if you choose private mode.

This repository also holds the tooling for FailEcho's own lab -- the fleet of test agents behind the scoreboard (failecho_fleet, failecho_agent, failecho_sandbox, deploy/). That tooling reads model-provider keys (GROQ_API_KEY, GEMINI_API_KEY, ...) and FailEcho's operator token from our servers' environment. You never need them, and nothing you install reads them. It is public so the lab's numbers can be checked, not because you run it.

app/web/static/vendor/ is Swagger UI, unmodified upstream build output with its checksums in its README; a test fails if it is ever edited.

Privacy

Privacy is a product feature, not a setting.

Collected — structured failure metadata only:

We do not want, and never store:

• prompts
• model messages
• tool arguments
• tool results
• request bodies and response bodies
• HTTP headers and cookies
• API keys, tokens and secrets
• customer names, emails and any user content
• credit-card data

Metadata only. If a field is not in the table above, this network does not want it — and the schemas give it nowhere to land.

How that is enforced:

1. The request schemas have no fields for any of it. Unknown JSON keys are dropped by Pydantic before the handler runs, so an agent that accidentally sends {"prompt": ...} cannot persist it here.
2. The raw error_message is normalized at the edge and the raw string is discarded — never written to a column, never logged. Only normalized_error survives.
3. Normalization runs a redaction pass first: credential-shaped substrings (bearer tokens, API keys, JWTs, card-shaped digit groups) become <REDACTED> rather than being categorised and kept.
4. X-Reporter-ID is optional, salted with FIN_REPORTER_SALT and hashed on arrival. The raw value is never stored. Rotating the salt makes existing hashes unlinkable.
5. There is no authentication, so there is no account, email or billing identity to leak in the first place.

Normalization examples:

Repository 918272 was not found              -> Repository <N> was not found
User [email protected] at 10.0.12.7 failed      -> User <EMAIL> at <IP> failed
GET https://api.example.com/v1/x?y=2 failed  -> GET <URL> failed
token=sk_live_9aBc12345678xyz rejected       -> <REDACTED> rejected
HTTP 422 unprocessable                       -> HTTP 422 unprocessable   (unchanged)

Small numbers survive on purpose: 422 and 500 are semantics, not identifiers. See app/core/normalize.py and app/core/privacy.py.

How the numbers are produced

Everything is deterministic arithmetic over observation counts. No model, no learned parameter, nothing you cannot recompute yourself.

Incident status (MVP heuristic, constants in app/core/config.py):

< 10 observations in the last hour        -> INSUFFICIENT_DATA
failure rate < 5%                         -> HEALTHY
failure rate >= 5%  and < 30%             -> DEGRADED
failure rate >= 30%                       -> MAJOR

The 5-minute window takes over from the 1-hour window once it holds at least 5 observations, so a fresh incident is not diluted by an hour of healthy history. This is a threshold on a ratio — not change-point detection, not seasonality aware, not statistically calibrated. It is labelled MVP logic on purpose.

Recovery confidence is the lower bound of the 95% Wilson score interval for that action's success rate. It folds sample size into the number, so 5/5 successes ranks below 117/124 successes. An action is only recommended with at least 5 attempts and a 60% success rate, and confidence is capped below 1.0. Thin evidence returns "recommendation": null. The network never fabricates confidence.

Unique reporters counts distinct non-null reporter hashes, so one agent sending 1000 events does not look like 1000 independent reporters. Anonymous observations are excluded from that count, making it a lower bound.

Abuse floor (V1)

No accounts, so the defences are structural rather than identity-based. Two independent layers, both transparent:

Per-reporter evidence cap. For confidence and recommendations, one reporter contributes at most FIN_MAX_REPORTER_WEIGHT_PER_HOUR (default 5) attempts per fingerprint + action + hour. Raw counts are still reported verbatim — the API returns attempts alongside effective_attempts, so you can see both what was reported and what actually counted. Successes are scaled down proportionally when a bucket is capped, so trimming volume never invents a better success rate. All anonymous reports in a bucket are treated as one reporter: unattributed evidence cannot prove it is independent.

Reporter diversity. A recommendation needs 5 effective attempts and a 60% success rate. Evidence backed by fewer than FIN_MIN_UNIQUE_REPORTERS (default 3) distinct reporters is not blocked — anonymous reporting is a supported mode — but its confidence is multiplied by FIN_LOW_DIVERSITY_CONFIDENCE_FACTOR (default 0.7).

Write rate limiting. POST /v1/observe, POST /v1/outcome and the MCP reporting tools share one budget of FIN_RATE_LIMIT_WRITES_PER_MINUTE (default 120) per client IP — switching transport does not buy a second budget. Reads are never rate limited; querying is the product. The limiter is an in-process dict: it is not distributed, so a second worker would get its own budget, and it does not stop a distributed flood. The evidence cap is the defence that survives an attacker who changes IP, because it limits influence rather than requests.

Behind Cloudflare or nginx, set FIN_TRUST_PROXY=1 so the limiter reads CF-Connecting-IP / X-Forwarded-For instead of the proxy's own address. Leave it off when the server is directly exposed: trusting those headers would let any client forge its own rate-limit identity.

Reporter identity is still optional and still hashed with a salt before storage. Raw identifiers are never written anywhere.

Retention and pruning

Raw observations are the hot path (the 5-minute and 1-hour windows read them directly) and also the thing that grows without bound. So:

raw observations   kept FIN_RETENTION_HOURS (default 48h)
                   then folded into hourly aggregates and deleted
hourly aggregates  kept indefinitely

Two aggregate tables: hourly_stats (successes, failures, unique reporters, latency sum/count per hour × service × operation × version × schema × source) and hourly_recovery_stats (attempts, successes, and the capped effective counts per hour × fingerprint × action).

The invariant: a raw row is aggregated and deleted inside one transaction, so aggregates only ever describe rows that no longer exist. "Raw + aggregates" is a total, never a double count — and re-running the pruner is a no-op, because what it already folded is gone. Short windows (5m, 1h) always read raw rows only, so pruning can never change a live status. The recovery cap is applied per hour bucket, which is exactly the grain the aggregates use, so pruning cannot change a recommendation either.

python scripts/prune.py                # use FIN_RETENTION_HOURS
python scripts/prune.py --hours 24     # override the window
python scripts/prune.py --dry-run      # report only, change nothing
python scripts/prune.py --vacuum       # also reclaim file space (briefly locks)

Retention window: 48h
Cutoff:           2026-09-08T09:51:18Z
Aggregated 18429 observations into 96 hourly buckets
Aggregated 812 recovery outcomes into 41 hourly buckets
Deleted 18429 raw observations
Deleted 812 raw recovery outcomes
Database size:    4.21 MB

Recommended cron (hourly, at :15) — not needed for local development:

15 * * * * /srv/failure-network/.venv/bin/python /srv/failure-network/scripts/prune.py >> /var/log/failure-network-prune.log 2>&1

Or use the bundled systemd timer: deploy/failure-network-prune.timer.

Project layout

app/
  main.py               FastAPI app, CORS, static homepage, /health, /llms.txt
  mcp_server.py         MCP tools + Streamable HTTP endpoint (same process)
  api/                  observe.py  query.py  outcome.py  services.py  deps.py
  core/                 normalize.py  fingerprint.py  intelligence.py
                        service.py  retention.py  ratelimit.py
                        privacy.py  config.py  clock.py
  db/                   database.py (async engine)  models.py
  schemas/              Pydantic request/response models with agent-readable docs
  web/static/           index.html  style.css  app.js   (no framework, no build)
                        logo.svg  favicon.svg  og-image.svg
client/
  failecho/             the public client package
    __init__.py         FailEcho: observe_tool_call, report_*, query
    adapters.py         ToolTelemetrySink -- the framework seam
    integrations/
      pydantic_ai.py    reference integration (optional dependency)
  failure_network.py    zero-dependency REST client (still supported)
  auto_recovery.py      the passive wrapper FailEcho is built on
  auto_recovery.py      failure-aware tool wrapper (reports + asks, never acts)
  example_agent.py      end-to-end REST usage example
examples/live_agent/
  tool_server.py        local tool that just shipped a breaking change
  tool_client.py        agent-side tool client with a stale cached schema
  network.py            MCP client (official SDK) for the four network tools
  agents.py             the autonomous loop: fail -> report -> ask -> recover
  run_demo.py           one command, six independent agents
scripts/
  seed_demo.py          synthetic demo telemetry (source='synthetic')
  prune.py              aggregate + delete expired raw rows
deploy/
  failure-network.service        systemd unit
  failure-network-prune.timer    hourly retention timer
  Caddyfile.failecho-dev         optional origin-level .dev redirect
LICENSE  SECURITY.md  CONTRIBUTING.md  .env.example
tests/                  the suite

app/core/service.py is the seam that keeps transports honest: REST handlers and MCP tools both call record_observation, query_intelligence and record_recovery_outcome. Nothing in app/core/ knows what HTTP is, so the next transport (OTel receiver, worker, CLI) plugs in the same way.

Demo data

scripts/seed_demo.py writes ~2000 observations and ~300 recovery outcomes across four services, every row tagged source='synthetic':

There are two kinds of non-real telemetry, and both are labelled at the row level by a source column:

demo_agent rows are real observations from real tool calls — the demo genuinely breaks a tool and genuinely recovers — but they are demonstrations, so they stay out of adoption metrics. Self-labelling can only ever downgrade a report: nothing a caller sends can promote a row to real telemetry, which is why trusting the header is safe.

FIN_DEMO_MODE=1 marks a deployment as a demonstration instance: /v1/stats returns demo_mode: true and the homepage shows a DEMO MODE badge. It never generates traffic — it only labels what is already stored. Nothing in this project fabricates telemetry at startup.

Both kinds are tracked separately everywhere they surface:

• /v1/stats reports real_observations_total, real_observations_24h, real_reporters_24h and real_failure_fingerprints excluding all demo rows, plus synthetic_observations and demo_agent_observations separately. They are never summed into one adoption number.
• the homepage renders real telemetry in the headline block and synthetic counters in a separate, visibly labelled block;
• /v1/recovery-intelligence flags every entry with demo_data: true|false (?include_demo=false hides them);
• POST /v1/query and the MCP check_tool_failure tool return demo_data_included, so an autonomous caller knows when it is acting on demo evidence.

Remove it all with python scripts/seed_demo.py --purge.

Configuration

Every setting is an environment variable; defaults are in app/core/config.py.

Deploying on a small VPS

Brand assets

app/web/static/logo.svg       the mark, inherits surrounding text colour
app/web/static/favicon.svg    the mark with fixed neutrals, for tab bars
app/web/static/og-image.svg   1200x630 social card

The mark is a failure event and its echo: one tall stroke in signal red, repeating outward and decaying. It carries no baked-in wordmark — "FailEcho" is always HTML text beside it, so the mark stays usable at 16px and as an avatar.

og-image.svg is served as-is. Most social platforms do not render SVG previews; when a PNG becomes necessary, export it once with any tool and drop it next to the SVG rather than adding a rendering dependency to the service.

Domains

failecho.com is the canonical public origin. Everything an agent or a human needs lives on it:

https://failecho.com/              homepage
https://failecho.com/mcp           MCP endpoint (Streamable HTTP)
https://failecho.com/docs          API reference
https://failecho.com/openapi.json  machine-readable schema
https://failecho.com/llms.txt      plain-text summary for agents

failecho.dev is a secondary domain and redirects permanently to failecho.com, preserving the path:

https://failecho.dev/*      ->  301  ->  https://failecho.com/*
https://failecho.dev/docs   ->  301  ->  https://failecho.com/docs
https://failecho.dev/mcp    ->  301  ->  https://failecho.com/mcp

Do this at the edge, not in the application. The app has no notion of a second domain and should not grow one.

www.failecho.com → failecho.com is handled at the origin by Caddy (redir https://failecho.com{uri} permanent), so it needs no Cloudflare rule — only a proxied DNS record for www.

Cloudflare (preferred) for the .dev domain. Add failecho.dev to the same account, then Rules → Redirect Rules → Create rule:

One rule covers both failecho.dev and www.failecho.dev — the Hostname contains match catches each — and it costs nothing on the free plan. Never serve a copy of the site from .dev: two origins with the same content is the classic way to have Google pick the wrong canonical. Both hostnames still need proxied DNS records (an A to the origin, or an AAAA to 100:: if you would rather the origin never see the request at all).

Caddy fallback, if you ever serve .dev from the origin instead — the config ships in deploy/Caddyfile.failecho-dev:

failecho.dev, www.failecho.dev {
	redir https://failecho.com{uri} permanent
}

api.failecho.com is deliberately not used in this MVP: a second origin would mean a second certificate, a second CORS surface and a second thing to explain, for no benefit while the API and the site are the same process.

Set the origin once, in one place:

FIN_PUBLIC_URL=https://failecho.com

It drives the canonical tag, Open Graph URLs, /llms.txt, the MCP endpoint shown on the homepage and every copyable example. No file in the codebase hardcodes the domain. Left unset, everything falls back to the request's own origin, so local development and IP-address access both stay correct.

Cloudflare checklist

DNS

www.failecho.com → failecho.com is handled by Caddy (redir ... permanent). The .dev hostnames are handled by the redirect rule above.

Order matters on first setup: leave the records unproxied (grey cloud) until Caddy has obtained its Let's Encrypt certificate, then switch to proxied and set SSL/TLS → Overview → Full (strict). Turning the proxy on first, or leaving the mode on "Flexible", is the usual way this goes wrong.

Caching. Never cache the live surfaces. Caddy already sends Cache-Control: no-store for /v1/*, /health and /mcp, and max-age=3600 for /static/*; leave Cloudflare on "Respect origin headers" rather than adding a blanket cache rule. Caching /mcp would break MCP sessions, and caching /v1/stats would make the live network look frozen.

Rate limiting. Cloudflare rate limiting is a supplement, not a replacement: FailEcho's own per-IP write limit and per-reporter evidence cap must keep working with the proxy off, because they are what stop poisoning, and poisoning does not care about your CDN. Nothing here requires a paid Cloudflare plan.

Deployment topology

Intended topology. The app binds to loopback only; TLS and the public address belong to Cloudflare and a local reverse proxy:

   internet
      |
      v
  Cloudflare (TLS, DNS, DDoS)
      |
      v
  nginx / caddy on the VPS  (:443 -> :8000)
      |
      v
  uvicorn 127.0.0.1:8000    FastAPI + SQLite (WAL)

Do not bind uvicorn to 0.0.0.0 in this topology. Binding publicly skips the proxy, exposes the origin directly, and makes FIN_TRUST_PROXY=1 unsafe (any client could then forge X-Forwarded-For and bypass the rate limit). Docker is optional and not required.

Step 1 — generate the reporter salt once and keep it.

sudo install -d -o failurenet -g failurenet /srv/failure-network/data
printf 'FIN_REPORTER_SALT=%s\n' "$(openssl rand -hex 16)" \
  | sudo tee /etc/failure-network.env > /dev/null
sudo chmod 600 /etc/failure-network.env

Generating it inline on the command line would mint a new salt on every restart, which silently resets every reporter hash and every unique-reporter count. Generate once, store once.

Step 2 — production command (what the systemd unit runs):

set -a; . /etc/failure-network.env; set +a
export FIN_DATABASE_URL="sqlite+aiosqlite:////srv/failure-network/data/failure_network.db"
export FIN_PUBLIC_URL="https://failecho.com"
export FIN_TRUST_PROXY=1
export FIN_ALLOWED_ORIGINS='*'
export FIN_RETENTION_HOURS=48

/srv/failure-network/.venv/bin/python -m uvicorn app.main:app \
  --host 127.0.0.1 --port 8000 --workers 1 \
  --proxy-headers --forwarded-allow-ips '127.0.0.1' --no-server-header

Note the four slashes in the SQLite URL: sqlite+aiosqlite:/// plus the absolute path /srv/.... Three slashes would make it relative to the working directory.

Step 3 — reverse proxy. Caddy:

failures.example.com {
    reverse_proxy 127.0.0.1:8000
}

nginx:

location / {
    proxy_pass         http://127.0.0.1:8000;
    proxy_set_header   Host $host;
    proxy_set_header   X-Forwarded-For $proxy_add_x_forwarded_for;
    proxy_set_header   X-Forwarded-Proto $scheme;
    proxy_buffering    off;   # keeps /mcp responsive
}

Then sudo cp deploy/failure-network.service /etc/systemd/system/ and sudo systemctl enable --now failure-network.

Environment checklist

Other notes

• One worker. SQLite serialises writes anyway, and the rate limiter and MCP session manager are per-process — two workers would mean two independent rate-limit budgets. Scale out only after moving to PostgreSQL.
• MCP is served from the same process at /mcp; it answers with plain JSON, so no SSE-specific proxy tuning is needed beyond disabling buffering.
• Retention: deploy/failure-network-prune.timer, or the cron line above.
• Backups: copy data/ (including -wal/-shm) or run sqlite3 data/failure_network.db ".backup backup.db". No downtime needed.

Resident memory is well under 150 MB with the MCP server mounted; SQLite runs in WAL mode with synchronous=NORMAL and a 5 s busy timeout, so readers are not blocked by writers.

Migrating to PostgreSQL later

Every column type is portable, timestamps are naive UTC, there are no SQLite-specific types and no expression indexes. Migration is FIN_DATABASE_URL=postgresql+asyncpg://... plus pip install asyncpg and one Alembic baseline.

Is it working?

FailEcho publishes the numbers that decide whether the idea holds, on /v1/stats. They are deliberately unflattering.

cross_agent_help_24h is the one that matters. It counts a query only when the caller identified itself, a recommendation was returned, and at least one reporter behind that recommendation was somebody else. Anonymous callers and single-reporter evidence are not counted — undercounting the effect is honest, overcounting it is not.

recovery_outcome_ratio_24h is the fragile one. Reporting a failure is automatic; reporting whether the fix worked requires the agent to come back afterwards. Without those reports FailEcho is an error counter.

Launch milestones

Internal experiment markers, not marketing claims:

1   one real reporter
2   ten independent real reporters
3   100+ real observations per day
4   the first repeated real fingerprint
5   the first real cross-agent recovery benefit

Milestone 5 is the hypothesis: an agent hits a failure, queries FailEcho, receives evidence generated by unrelated agents, changes behaviour, and recovers. Everything before it is plumbing.

MVP limitations

Stated plainly, because pretending otherwise would make the network less useful:

• No authentication. Anyone can report anything. The per-reporter evidence cap and rate limiter raise the cost of poisoning the statistics; they do not make it impossible, and a distributed flood from many IPs would still get through.
• No reputation scoring. Reporters are counted, not ranked. A reporter that has been right a thousand times counts the same as a fresh one.
• The rate limiter is in-process and not distributed. One uvicorn worker, one budget. It resets on restart.
• No sophisticated anomaly detection. Status is a fixed threshold on a failure ratio over two fixed windows.
• Unique-reporter counts in aggregates are lower bounds. Reporter identities are not retained past pruning, so merged buckets keep the maximum per-bucket count rather than a true distinct count.
• No OpenTelemetry ingestion yet, no TypeScript SDK yet, no payments (x402 or otherwise). Everything is free.
• SQLite is a prototype-stage choice. Retention keeps the file small, but a busy network will eventually want PostgreSQL (a URL swap plus asyncpg).
• Recovery actions are free-form strings, so refresh_schema and refreshSchema would be counted separately if agents disagree on spelling (input is lowercased and space-normalized, which handles the common cases).

Evidence

• docs/claims.md — every claim FailEcho makes, the exact evidence behind it, and the ones it must not make
• lab.failecho.com/fleet — the live scoreboard, losing groups included

Contact

• General: [email protected]
• Integration help: [email protected]
• Security reports: [email protected] — see SECURITY.md; please do not open a public issue for a vulnerability

Licence

MIT.