AI agent failure diagnosis & root-cause debugging

Snap your agents back to a working state

When an AI agent breaks — an infinite loop, a hallucinated tool schema, a burned token budget — Snapback diagnoses the root cause in milliseconds and hands back the fix. AgentOps for the self-healing era, for developers and the agents themselves.

Free to start (2,000 diagnoses) · then ~$0.0032/diagnosis via x402, or $99/mo for teams · no card to try

an agent queries snapback mid-failure
GET /v1/solutions?error=ModuleNotFoundError&pkg=vllm

{ "problem": "vllm import fails under Python 3.13",
  "solution": "pin vllm==0.6.3; avoid 3.13 wheels",
  "confidence": 0.94 }

Pay per call with x402 — no account needed. Discover tools via MCP.

what's a snapback?

A Snapback is your agent's autopsy report. You get three things: what broke, why it broke, and how to fix it.

Technically: a verdict on a failed agent run — the failure class, the exact step and agent that broke, the root cause, and the fix, plus what it cost in tokens. You send a trace, you get a Snapback back. Most come straight from the pattern library in milliseconds with no LLM call; the rest are diagnosed live and fold back into the library, so every failure makes the next diagnosis faster. It's not a wall of logs — it's the answer.

see it work

From cryptic error to fix, in 30 seconds

1 · YOUR AGENT HITS AN ERROR
unable to get local issuer certificate
2 · ONE CALL · FREE · <150MS · NO LLM
diagnose_infra_error({error})
3 · YOU GET THE FIX
dns-tls family · confidence 1.0 · source: library
The server is sending an incomplete cert chain — serve the intermediate certificate, not just the leaf.

The fix the base model usually gets wrong — instant, from a curated library.

See all 58 guides →

The whole failure lifecycle

Most tools show you the post-mortem. Snapback covers before, during, and after.

Pre-flight

Catch it before it ships

Smoke-screen your agent's tool schemas and prompts against thousands of known failure patterns — before a single token is spent.

  • Validate tool schemas
  • Test for prompt drift
  • Flag known-bad patterns
During

Diagnose the moment it breaks

Snapback intercepts the failure live, classifies the root cause, and tells you exactly which agent, which step, and why — no scrolling ten thousand lines of logs.

  • Catch loops & timeouts
  • Root-cause classification
  • Which agent, which step
After

Teach it to self-correct

The fix flows straight back to your agent's memory files. Mark whether it worked, and the registry sharpens for every agent that hits the same wall next.

  • Fix → agent memory (soul.md)
  • Did it work? 👍 / 👎
  • The registry compounds
Browse the failure library →

Every failure class Snapback diagnoses — the catalog behind the pattern library.

For developers

Stop babysitting agents. Get a structured verdict — what failed, where, and the fix — instead of reading ten thousand lines of terminal output after the token budget is already gone.

For agents

A machine-readable registry of defeats and workarounds. Query it mid-failure over MCP or HTTP, pay per call with x402, inject the fix, and skip the three-wrong-guesses loop entirely.

Built for local-first agents

Made for the OpenClaw & Hermes crowd

If you run OpenClaw on disk or Hermes through Ollama or vLLM, your agent already writes its memory and skills to local markdown. Snapback reads those failure signals, diagnoses the crash, and patches the workaround straight into soul.md — so the next reasoning turn already knows the fix.

  • Self-hostable — telemetry never leaves your machine
  • Local-first daemon watches your agent's activity stream
  • Open-source core (MIT) — inspect every line

clawed it back

FOR AI AGENTS

Building with an agent?
It can figure us out itself.

Two doorways — one for reading, one for calling. Both free to discover, no signup.

READpoint an agent here
curl snapback.sh/llms.txt

A plain-text map of what Snapback is, the tools, the trace format, and pricing. Your agent reads it and works out the rest.

CALLpoint an MCP client here
api.snapback.sh/mcp

The same catalogue as callable MCP tools — discover, diagnose, catch loops mid-run, and (optionally) pay per call with no account.

Production path: point an MCP client at api.snapback.sh/mcp and pay per call via x402 — no account, no monthly limits. Just validating the integration? An agent can self-register for a starter allowance first (POST /v1/agents/register).

x402 accepts USDC on Solana and the EVM chains Base · Arbitrum · Polygon · Avalanche — your agent's wallet pays on whichever chain it holds USDC. Same price, same verdict.

Discovery, docs search, and mid-run loop checks are always free. See the docs or pricing.