Prove what your AI agents did — without storing what they saw.
Agent Evidence Recorder turns every critical action an agent takes — model calls, tool calls, human approvals, redactions and errors — into a chronological, hash-anchored evidence packet a compliance, legal or security reviewer can trust.
- Provable
- A chronological timeline of every model call, tool call, approval and error.
- Redacted by default
- Sensitive values are never stored in the clear — only hashes and findings.
- Portable
- One JSON audit packet, anchored by a sha256 seal, reviewable outside the system.
What's new · collector-js 0.7
Verified human approval
The agent asks; only a signed-in user of your workspace approves. The approval is bound to the exact action and parameters, and every step lands in the evidence chain.
run.requestApproval()
Signed receipts, offline verification
Sealing a run returns a receipt. Keep it, and anyone can later check an exported packet against it — no network, no AER account.
npx @ktlsr/aer-verify
Delivery that survives outages
Retries honour Retry-After, events queue while AER is unreachable, and an optional file store keeps them across restarts.
recorder.shutdown()
Already using AER? npm i @ktlsr/collector-js@latest — older SDK versions keep working unchanged. Upgrade notes
Stored redacted — raw sensitive values are never persisted. Evidence packets are hash-anchored.