Mantyl is the handover layer for agentic development. It reads your repository and your coding agent’s session history, executes real verification in a sandbox, and produces a project passport that a recipient can independently check. Everything runs on your machine and nothing is uploaded unless you explicitly publish.
Three things make the difference between a thin passport and a full one. Set them up before your first run and the first result will be the real result.
| Requirement | Why it matters |
|---|---|
| Node 20.12+ | Runs the CLI itself. Check with node --version. |
| Docker Desktop | Powers the verification sandbox: your project’s install, build and tests execute in an isolated container, which is where the passport’s strongest evidence comes from. Make sure it is installed and running before mantyl verify. Without it, checks are honestly recorded as skipped and the passport carries far less proof. Get it at docker.com. |
| Claude Code | The agent history is where claims and decisions come from. Run mantyl from the same project folder you used with your agent so it finds the sessions. Cursor chat history (on Node 22.13 or newer) and OpenAI Codex CLI sessions are also read, in beta. The supported agents page covers each store in detail. Projects without agent history still work, with a repository-only passport. |
Mantyl ships as a single npm package.
Run mantyl doctor first, every time. It tells you exactly which of the prerequisites are ready on your machine and what to do about any that are not.
Run these four commands from the root of a project you want to hand over. Each one is useful on its own and none of them talks to a server.
Artefacts land in the .mantyl directory. The canonical contract is passport.json, stamped with a digest that binds it to the exact commit and file contents it describes.passport.html is the same content as a readable report.
Every statement in a passport carries a truth status and a source reference. The repository is evidence. Executed checks are evidence. What an agent said it did is recorded as exactly that, a report, and when the repository contradicts it the passport says so plainly. The full vocabulary is on the truth model page.
Session transcripts pass through fail-closed secret redaction before they are stored or analysed. If you enable the optional LLM analysis, it sees redacted excerpts only and everything it writes is permanently marked as inference.
Mantyl reads Claude Code session stores, plus Cursor chat history and OpenAI Codex CLI sessions in beta, with more agentic coding tools planned. From the transcripts it extracts claims, which are statements about work done, and decisions, which are choices made along the way. Your own recorded words become creator confirmed entries, because the person who made the call said so on the record.