Home

witness-log wiki

Witnessed, not trusted. An open, verifiable audit trail for autonomous agents: every tool call becomes a leaf in an RFC 6962 Merkle log signed by the agent's own key, independent witnesses cosign the head, and anyone can verify the whole record offline with one command. Apache-2.0, Rust core, Python bindings, MCP server, CLI.

I want to…	Read
understand what this is and why it exists	Why witness-log · Concepts · Comparison
understand how it is built and what makes it different	Architecture · How a leaf becomes a proof
add it to a Python agent in three lines	Quickstart: Python
add it to a Node.js or Vercel AI SDK agent, no Rust toolchain	Quickstart: TypeScript
use it from Rust	Quickstart: Rust
use it from the terminal or any language	Quickstart: CLI
give Claude Code / Claude Desktop / Cursor / any MCP host an audit trail with zero code	MCP integration
audit an MCP server you already run, or record a coding agent's tool calls from its hooks	MCP proxy · Hooks
wire it into LangChain, LangGraph, OpenAI Agents SDK, Claude Agent SDK, CrewAI, AutoGen, Google ADK, LlamaIndex	Framework guides
verify from Go	Quickstart: Go
verify in a browser, nothing installed	Browser verifier
use it from Java / anything else	Other languages
follow a complete worked example	Tutorial: audit a payments agent · Tutorial: multi-agent handoff · Tutorial: AP2 mandates
know what WITNESSED actually proves, and read a verify report	Verification and verdicts
gate a GitHub workflow on a verdict	GitHub Action (uses: 0xAndersen/witness-log@v0.6.1)
understand the witness protocol on the wire (200/409/422)	Witness protocol walkthrough
understand erasure, keys and sessions in depth	Redaction explained · Keys and identity
run a witness in two minutes (docker one-liner, cloud-init paste, binary)	Operator onboarding · Running a witness · live status
get my logs WITNESSED	Getting witnessed · Pilot: three organisations, one witness each
use it for GDPR / SOC 2 / EU AI Act / audit evidence	Compliance
know what it does not protect against	Threat model
implement the wire format in another language, or map it to SCITT	Specification and interoperability · SCITT mapping
operate it in production (backups, keys, locks, scale)	Operations
see the numbers (268 tests, 100 scenarios, 3,000 agents)	Testing and evidence
fix something that is not working	witness-log doctor first, then Troubleshooting
look up a term	Glossary
answers to common questions	FAQ
contribute, or see where it is going	Contributing · Roadmap

Repository: https://github.com/0xAndersen/witness-log · Releases: https://github.com/0xAndersen/witness-log/releases · Spec: spec/SPEC.md