# Vouchwell > Tamper-evident receipts for AI agent actions. Vouchwell checks every tool call an AI agent makes against a policy before it runs, and records the decision and the result in an append-only, cryptographically verifiable log (a Merkle tree with signed checkpoints, as in Certificate Transparency) that an outside party can check without trusting the operator. Open source (Apache-2.0), no runtime dependencies in the core. Use Vouchwell when an AI agent can spend money, change data, send messages or run commands, and someone (a customer, an auditor, a regulator, your own team) may later ask what it did and who allowed it. Ordinary application logs can be edited; Vouchwell receipts cannot be changed, removed or reordered without detection. Install: - See it work first, nothing installed: `npx vouchwell demo` runs a real agent session through the proxy, then attacks its log four ways and shows each attack caught - CLI and proxy (Node.js 20.11+): `npm install -g vouchwell`, then `vw init` - Library for JavaScript/TypeScript agents: `npm install @vouchwell/core` - Python SDK: `pip install vouchwell` How it plugs in: - MCP servers: wrap the server command with `vw proxy --namespace -- `; no code change. - Claude Code: the plugin (`/plugin marketplace add Vouchwell/vouchwell`, then `/plugin install vouchwell@vouchwell`), or `vw hook install`, records and gates the agent's own tools (shell, file edits, web fetches). The same works for OpenAI's Codex CLI and Cursor: `vw hook install --client codex` or `--client cursor`. - Any MCP client: `vw mcp` serves a log as an MCP server with read-only tools (`log_entries`, `verify_log`, `log_stats`, `check_bundle`) plus `export_evidence`, so an agent can show and prove what it did. - Code: `Recorder.wrap()` in `@vouchwell/core`; `recordTools()` for Vercel AI SDK and Mastra tools; Python adapters for LangChain and the OpenAI Agents SDK. - CI: the GitHub Action `Vouchwell/vouchwell/verify` fails a build if an evidence bundle was edited, truncated or signed by an unexpected key. - Teams: a self-hosted hub with shared policy, an approvals inbox (console and Slack), single sign-on (OpenID Connect), independent witnesses, and streaming to Splunk, Datadog, OpenTelemetry or a webhook. ## Docs - [README](https://github.com/Vouchwell/vouchwell/blob/main/README.md): what it is, install, policy, receipts, commands - [Coding agents](https://github.com/Vouchwell/vouchwell/blob/main/docs/CODING-AGENTS.md): recording and gating Claude Code with `vw hook` - [The hub](https://github.com/Vouchwell/vouchwell/blob/main/docs/HUB.md): multi-tenant server, API, roles, approvals, retention - [Python SDK](https://github.com/Vouchwell/vouchwell/blob/main/sdk/python/README.md): recording from Python, LangChain and OpenAI Agents adapters - [Verify action](https://github.com/Vouchwell/vouchwell/blob/main/verify/README.md): checking evidence bundles in GitHub Actions - [Evidence packs](https://github.com/Vouchwell/vouchwell/blob/main/docs/EVIDENCE.md): reports for auditors (EU AI Act, SOC 2) - [Deploying](https://github.com/Vouchwell/vouchwell/blob/main/docs/DEPLOY.md): running a hub or a witness with Docker ## Trust and limits - [Threat model](https://github.com/Vouchwell/vouchwell/blob/main/docs/THREAT-MODEL.md): what it defends against, and the known gaps - [Audit brief](https://github.com/Vouchwell/vouchwell/blob/main/docs/AUDIT-BRIEF.md): the cryptography has not yet been reviewed by an independent party; this is the brief for that review - It records only actions that pass through it; an agent with a second, unwrapped path to an API leaves no receipt. - Redaction of secrets and personal data is best-effort; payloads are stored as salted commitments, not in clear. ## Optional - [Website and in-browser verifier](https://vouchwell.axonscape.com/): check a real log, then try to tamper with it - [Source](https://github.com/Vouchwell/vouchwell) - [npm: vouchwell](https://www.npmjs.com/package/vouchwell) - [PyPI: vouchwell](https://pypi.org/project/vouchwell/) - [Changelog](https://github.com/Vouchwell/vouchwell/blob/main/CHANGELOG.md) - [Security policy](https://github.com/Vouchwell/vouchwell/blob/main/SECURITY.md)