agentwitness 0.0.1

Verifiable evidence for AI-assisted engineering

agentwitness

Verifiable evidence for AI-assisted engineering.

Capture what your coding agent did, under whose authority, with what scope — and export signed evidence for PR review, incident response, security questionnaires, and AI governance audits.

Status

Pre-release. Version 0.0.1 reserves the package name; the v0 implementation is in active development.

• Repository: https://github.com/ConceptPending/agentwitness
• Homepage: https://agentwitness.dev

Scope

agentwitness captures agent actions inside Claude Code via the platform's hook system, signs them with an Ed25519 key under the operator's control, chains them with a verifiable hash, and produces evidence bundles that auditors, security reviewers, and third parties can independently verify.

Four ways the same artifact gets used:

1. Engineering — what changed and why?
2. Security — what was the agent allowed to do, and did it stay in scope?
3. Compliance — what evidence can we show?
4. Legal / procurement — can we verify contractor or vendor AI usage?

What this proves — and what it doesn't

A signed, hash-chained event log proves that captured events were not modified after capture, and that delegated actions trace back to a root signing key.

It does not by itself prove completeness: in a hostile local environment hooks can be disabled, files deleted, and tools run outside the instrumented path. Completeness is a process problem — approved plugin configuration, CI enforcement, retained checkpoints — not a cryptography problem.

A fuller threat-model document ships alongside the v0 release.

License

Apache License 2.0 for code. The accompanying specification will be licensed CC-BY-4.0 in the spec/ directory of the v0 release.

Author

Nick Williamson — https://nickw.info