If Claude and Codex both pick up the authentication change, separate worktrees will let both spend time on it. You can end up reviewing two versions of the same work.
A shared claim gives the second agent a chance to pick something else before it starts.
That's the problem I built Kith around. Agents use it through MCP to claim a task and file scope, work in a separate Git worktree, and coordinate resources on the same machine. It supports Claude Code, Codex and Cursor CLI.
For example, an agent claims src/auth/**. A second agent requesting an overlapping exclusive claim gets a conflict response. They can also reserve different ports or use a named mutex for a command that needs to run one at a time.
When an agent finishes, Kith prepares integration in a staging worktree and runs the verification command you supply. A failed check leaves the target unchanged and keeps the source worktree for repair. Choose a check that tests your project: without one, Kith checks Git mergeability only.
It's open source and runs locally. Kith itself needs no hosted service or model API key.
The released package is kith-coord:
npm install --global kith-coord@0.1.1 cd /absolute/path/to/your/project kith init --dry-run kith init Review and commit the generated configuration, restart your coding client, and approve its MCP connection. Give each parallel agent a distinct connection, ID and worktree. Keep the connection alive and heartbeat before session expiry.
Full supervised-command workflows require macOS/Linux or WSL. Native Windows has partial MCP support. Claims coordinate cooperating agents; they aren't a filesystem sandbox.
The repository has the setup instructions, client validation and recovery boundaries.
Disclosure: Drafted by an AI agent using Kith's implementation and recorded validation.
