writ

Lets an agent request a commit-time policy decision (ALLOW, DENY, or STEP_UP) before a consequential write, with a tamper-evident audit trail.

What it does

Writ is a gate an MCP-compatible agent calls before performing a consequential write. The agent calls writ_check with a sponsor, agent, verb, target, and purpose, and receives ALLOW, DENY, or STEP_UP; an ALLOW returns a 90-second token bound to that exact write, which writ_verify_token re-checks immediately before execution. Sponsor-side tools grant a STEP_UP, revoke or reinstate a principal, and agent-side tools read receipts and tenant verb policy. A writ_sandbox tool offers a free demo grant without a key.

When to use it

Use it when an assistant performs writes that should be authorized at the point of action rather than only at setup, and when you want a decision record for each attempt. It suits teams that need a human approval step for sensitive operations and an audit trail of what was allowed or denied.

Requirements

Runs locally over stdio as the writ-mcp command from the PyPI package writ-mcp, requiring Python 3.9+ (installable with uvx). A Writ API key is required in the WRIT_API_KEY environment variable; sponsor tools additionally need WRIT_SPONSOR_TOKEN. WRIT_BASE_URL optionally overrides the default API endpoint, so network access to the Writ API is needed.