writ-mcp

The Writ gate as an MCP server. Give any MCP-compatible agent point-of-action control: the agent calls Writ before a consequential write, gets back ALLOW, DENY, or STEP_UP, and every outcome creates an audit receipt.

Install

pip install writ-mcp

Requires Python 3.9+. Installs the writ-mcp command (stdio transport).

Configure

export WRIT_API_KEY="writ_..."        # required; get one free: POST /v1/keys with an email
export WRIT_SPONSOR_TOKEN="..."       # only for sponsor tools (grant, revoke, reinstate)
# export WRIT_BASE_URL="..."          # default: https://api.withwrit.com

Add to your agent

Claude Code:

claude mcp add writ --env WRIT_API_KEY="$WRIT_API_KEY" -- writ-mcp

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "writ": {
      "command": "writ-mcp",
      "env": { "WRIT_API_KEY": "writ_..." }
    }
  }
}

Any MCP client (generic stdio config):

{
  "command": "writ-mcp",
  "env": {
    "WRIT_API_KEY": "writ_...",
    "WRIT_SPONSOR_TOKEN": "writ_sp_..."
  }
}

Hermes (NousResearch/hermes-agent) catalog entry — once published, this package is installable from the optional-mcps catalog:

hermes mcp install writ

The check-before-write loop

The tool descriptions teach the agent this flow, but the short version:

1. writ_check(sponsor_id, agent_id, verb, target, purpose) — before the write.
2. ALLOW → you get a 90-second authToken bound to that exact write.
3. writ_verify_token(auth_token, verb, target, purpose) — immediately before executing, to prove the authorization still matches what you're doing.
4. DENY → do not proceed. STEP_UP → a human sponsor approves (via writ_grant or the dashboard), then check again.

Tools

Try it with no key: writ_sandbox → writ_check with verb="demo_write".

Source

Public repo: withwrit/pywrit (mcp/ directory). License: MIT.