RECOVER

A failed machine action happened. What should happen next?

RECOVER is a tiny deterministic decision API for AI agents, MCP clients, automations, and ordinary backend systems. It converts failure context into one of six actions:

RETRY_NOW · WAIT · REPAIR_REQUEST · VERIFY_FIRST · ESCALATE · ABORT

Use it immediately after a failed API/tool/action and before an automatic retry—especially for side-effecting operations, ambiguous timeouts, rate limits, malformed requests, exhausted retry budgets, and uncertain execution outcomes.

Live production

Base URL: https://recover-api-production-bbba.up.railway.app

MCP Streamable HTTP endpoint: https://recover-api-production-bbba.up.railway.app/mcp

Activity dashboard: https://recover-api-production-bbba.up.railway.app/activity

Safe demo: https://recover-api-production-bbba.up.railway.app/demo

Related MCP server: ErrorLens MCP

REST

POST https://recover-api-production-bbba.up.railway.app/v1/recover

{
  "method":"POST",
  "status":503,
  "error":"upstream timeout",
  "attempt":1,
  "sideEffect":true,
  "idempotencyKey":false
}

Example response:

{
  "decision":"VERIFY_FIRST",
  "retry":false,
  "risk":"high",
  "reason":"A side-effecting non-idempotent action may have completed before the failure became visible.",
  "next":"Verify external state or obtain an idempotency key before retrying."
}

MCP

Tool: recover_decide

Point an MCP client at:

https://recover-api-production-bbba.up.railway.app/mcp

The server is designed for machine discovery and includes server.json metadata for the official MCP Registry.

Why call RECOVER?

Retry libraries answer how to retry. RECOVER answers the more important question first: what is the safest next action after this specific failure?

Good call sites include API clients, AI agents, browser automations, workflow engines, MCP hosts, job runners, payment/order flows, and any system where blindly retrying can duplicate a side effect or amplify a failure.

Analytics

GET /activity displays durable, privacy-safe usage telemetry and a permanent FIRST 10 REAL-WORLD STRANGER CONVERSIONS scoreboard. Real-use events are stored separately so discovery traffic cannot push milestones out of history. Raw IPs and request payloads are not stored.

Deploy

The repo is Railway-ready with Dockerfile and railway.toml. Mount a persistent volume at /data so the permanent real-use ledger survives restarts.