Skip to main content
Agents now write to real systems: orders, payments, tickets, email. When a run goes wrong, teams need three answers: what changed, why, and how do we undo it? Reversi sits between an agent and its tools and gives you:

Tool manifests

Declare per tool: how reversible it is, what undoes it, which permission it needs. Versioned, reviewable, signable.

Saga rollback

Undo a run newest-first, idempotently, with retries and read-backs, and never claim an irreversible step was undone.

Hash-chained ledger

Who did what and why. Edits, deletions and reordering are detectable. Credentials never reach it.

Policy and approvals

Rules as code, rate limits, content checks, and human approval bound to the exact call, with quorum, delegation and break-glass.

MCP gateway

Put it in front of any MCP server. Starts in observe mode with no setup; journal, enforce, approve and roll back without changing the agent.

Console

See runs, decide approvals, run rollbacks, edit manifests and policy. Off by default.

Simulate before you ship

Replay captured calls through a candidate policy and see what would change.

Signed evidence

Ed25519-signed bundles, auditor reports and externally anchored ledger heads.

​
What is in the box

Start with the quickstart, or go straight to the gateway if you already run MCP servers. To run it for a team, see GitOps and deployment.
Reversi records and enforces what manifests declare. It cannot discover what a tool really does. A manifest that lies makes Reversi lie. Many parts (the connectors, notification channels, framework adapters, Docker and Helm files) are built and tested against fakes but not against real services or clusters. Read limits, the roadmap and the threat model.