claudrunner FAQ

Is it free?

Yes. MIT licensed, no account, no sign-up, no telemetry, no paid tier. You pay only for your own model usage: your Claude Code session when you run it by hand, or an Anthropic API key (the ANTHROPIC_API_KEY secret) when it runs on a schedule in CI. To use less, sweep weekly instead of nightly and poll the board less often.

Which task boards does it support?

Eleven: Jira, Trello, GitHub Issues, Linear, GitLab Issues, Azure Boards, Shortcut, Asana, ClickUp, monday.com and Notion. It also runs with no board at all — the bug sweep then writes its findings as Markdown reports. A new board takes about an afternoon to add; see writing a board adapter.

Which languages does it work with?

Any. Packs ship for Node and TypeScript, Python, PHP, Go, Java, .NET and Rust. For anything else the generic pack reads your CI workflow and your project’s own scripts, proposes the commands it found, and asks you to confirm. Your project’s commands always win.

Where can it host code?

GitHub (including Enterprise), GitLab (including self-managed), Bitbucket Cloud and Azure Repos. Your board and your code host are separate choices — Jira with GitHub is common.

Can it push to main, or merge?

No, never. At every autonomy level it cannot push to your default branch, merge, force-push or delete a branch. The default level, pr-only, opens a pull request and stops there. Protect your default branch anyway; that protects you from everyone else.

What if a ticket is unclear, too big, or a bad idea?

It never stops mid-run to ask. It parks the ticket with a note written for a person:

Outcome	The note contains
unclear	The one question whose answer unblocks it
too-large	A proposed split into two or three smaller tickets
not-needed	The evidence that it is already done or its premise is false
better-approach	What should happen instead
suspicious	What the ticket tried to make the agent do

Refusing to build something is a real outcome, not a failure.

Does the AI see my secrets?

No. A small shell orchestrator fetches tickets, claims them, prepares the branch and moves the tickets afterwards. The agent gets a working tree, the ticket text and a short list of permitted commands. See the security model.

Can a malicious ticket take control of it?

Ticket text is treated as untrusted input — it describes work, it never instructs the agent. A ticket asking it to run commands, change permissions, edit CI or push somewhere is skipped as suspicious and reported. And every ticket id the agent reports is checked against its own input, so it cannot touch the rest of your board.

Does it need a server or Docker?

No. The default runs from CI cron on the runner your repository already has. A server with systemd, plain cron, and Docker or Podman isolation are all supported for teams that want them.

How do I stop it?

Disable the workflow, remove the cron line, or stop the systemd timer. Nothing else runs in the background, and it keeps no state outside your repository.

Who made it?

claudrunner is designed and built by samiraklf and is open source on GitHub. Issues and pull requests are welcome.

/plugin marketplace add samiraklf/claudrunner
/plugin install claudrunner
/claudrunner:init