You open a pull request. CI runs. Something fails. You fix it, push again. A reviewer leaves comments. You address them, push again. CI passes. You wait for approval. You merge. Repeat for every PR.
This loop is mechanical. It doesn't require creativity - it requires persistence. And that's exactly what Claude Task Master handles.
claudetm merge-pr 147
That command owns the entire merge loop:
The tool doesn't just retry - it understands the codebase, runs tests, and makes targeted fixes.
This repository's own PRs #147 through #150 were merged using claudetm merge-pr. The maintainer didn't babysit any of them.
merge-pr is one entry point. The full autonomous workflow starts with:
claudetm start "Add user authentication with tests"
That command:
If you run multiple Claude subscriptions, profiles keep them isolated:
claudetm --profile team-a start "..."
claudetm --profile ci-bot --api-key $KEY start "..."
Each profile gets its own ~/.claudetm/profiles/<name>/ directory so credentials never clobber each other.
The REST API, MCP server, and HMAC-signed webhooks expose the same lifecycle programmatically. A dashboard, a bot, or a CI pipeline can dispatch tasks and react to progress without running the CLI interactively.
uv tool install claude-task-master
claude
/login
cd your-project
claudetm start "Your goal here"
Requires Python >= 3.10, authenticated Claude CLI, and authenticated GitHub CLI.
Open source, MIT licensed. https://github.com/developerz-ai/claude-task-master