Debug your AI agents like you debug your services.
Every decision, replayable.
Every tool call your agents make (Claude Code, Cursor, Cline, Codex, Continue) lands in one signed chain. When something breaks in production, open the session, replay the exact sequence of decisions, and see the context each agent held. When a reviewer asks later, the same record verifies offline against your pinned public keys.
Start a free pilot → Run the notebook (10 min)
What you get in the first 10 minutes
Wire once
One EtchLangChainCallback or EtchAnthropic
wrapper. Zero changes to your agent logic. Zero perf overhead
at critical path (async recording).
See every decision
Sessions dashboard groups every conversation. Timeline swimlane, provenance DAG, and full text of every tool call are one click away.
Prove nothing was rewritten
Every 100 events, Etch closes an epoch and signs the Merkle root with a hybrid Ed25519 + SLH-DSA envelope. Auditors verify offline against your pinned public keys.
Use cases we hear from VPs of Engineering
1. Pre-production release gate
Run your candidate agent through your regression suite. Every step lands in Etch. Compare against the current production agent using Shadow Mode. If any divergence exceeds your threshold, block the release. Signed evidence for your CTO of what you tested and why you shipped.
2. Post-incident forensics
Customer reports the agent misfired at 10:23:47. Open the Sessions tab → click Replay on the flagged event → see every prior tool call, every retrieved context, every LLM decision that led there. Iterate the fix. Prove to the customer the fix works with the same substrate.
3. Regulator inquiry
"Prove the model didn't use protected demographic attributes to make this loan decision." Open the incident replay for the flagged decision, download the signed bundle. Regulator verifies offline against your public keys. Case closed.
4. Multi-agent team observability
Your team uses Claude Code, Cursor, Cline, Codex, and Continue in different combinations. Every developer's activity attributes to their git email. Team tab shows per-developer rollups. Feature tags link cross-agent work on the same feature.
The 3-pillar proof kit
Everything that lets you trust Etch is verifiable from git clone:
1. Trust assets
Dogfooding report from 62 hours of real production, 152k recorded events, 19k signed epochs. Adversarial suite with 13 tests, all passing. Compliance mapping to EU AI Act, SEC 17a-4, NIST AI RMF, and FIPS. Crypto correctness suite with 28 tests including RFC 8032 test vectors.
2. Runnable notebook
LangChain agent + 10 evaluation scenarios + LLM-as-judge + signed Evaluation Scorecard. 10 minutes from clone to cryptographically-verifiable eval bundle.
3. Integration docs
Anthropic, OpenAI (+ any OpenAI-compat provider: Groq, OpenRouter, HuggingFace router, etc.), LangChain callback, LangGraph shim, MCP proxy, OTel receiver. Any framework, any model.
Why cryptography, not just observability
| Question | Observability (Langfuse / Arize) | Etch |
|---|---|---|
| Can you query what the agent did last week? | Yes | Yes |
| Can you prove the log wasn't silently rewritten? | No, trust the vendor's DB | Yes. Merkle root + hybrid signature |
| Can an auditor verify without contacting the vendor? | No | Yes, pinned public keys, offline verify |
| Does the record survive a post-quantum break of Ed25519? | N/A | Yes, hybrid with SLH-DSA-SHA2-128f (FIPS 205) |
| Can the vendor be compelled to modify records under duress? | Yes. DB is theirs | Not without breaking the chain visibly |
Observability solves "what happened?" for your engineers. Etch solves "can we prove it?" for your customers, your regulators, and your insurance carrier.
Ready to try it?
Sign up for a pilot, free during beta → Skip signup, run the notebook first →
Etch is bootstrapped, based in Bangalore, and OSS-first (source on GitHub). Reply to your signup email if you want a walkthrough, typically a 20-minute call, no sales team.