Install & setup

Grab the desktop app from Download and run the installer. It sets up the strazh CLI, the local MCP server, and the dispatch console together — you don’t need a separate CLI install.

Confirm the CLI landed (open a new terminal after installing):

Strazh only dispatches into repos you’ve explicitly registered. Add every working directory you actually want a coding agent to touch:

Anything outside this list is rejected before a process ever spawns: see it happen on the homepage. There is no bypass route; the gate checks every dispatch against this list, every time. This is the same allow-list file the desktop app reads, so registering a repo here also registers it there.

Point Strazh at an agent and a registered directory:

Strazh launches the agent itself, so it is the one thing that can stop it. Environment scrubbing keeps secrets out of the agent’s process, and a timeout tree-kills the whole process group if a dispatch runs away. Every dispatch, approved or rejected, is written to a local, queryable audit log (~/.strazh/audit.jsonl).

Prefer to dispatch from inside Cursor, Codex CLI, or another MCP client instead of the CLI directly? Strazh exposes the same gate as an MCP server: a singlePOST /mcp JSON-RPC endpoint on the same local server the desktop app runs (start it with python -m strazh.server if you’re not running the desktop app). Point your client at it with the local token the server prints on startup:

This is a local HTTP endpoint, not a separate stdio-based process. A client that only speaks stdio (Claude Desktop, as of this writing) needs a stdio↔HTTP bridge such as mcp-remote in its config to reach it. Clients that support HTTP-based MCP servers directly (Cursor, Codex CLI) can point at the URL above with no bridge. Either way, a dispatch triggered by an MCP client goes through the exact same allow-list, cap, and audit trail as one triggered from the CLI or the desktop app: no separate, less-guarded path.