claudrunner compared with other ways to run an AI coding agent

There are several ways to put an AI coding agent to work on a repository. This page is an honest comparison, so you can pick the one that fits.

At a glance

	By hand in a terminal	One CI workflow per event	Hosted coding agent	claudrunner
Picks up tickets by itself	❌	Only the event it is wired to	Usually, from its own queue	✅ From 11 task boards
Hunts bugs you did not report	❌	❌	Varies	✅ Scheduled sweep with evidence
Independent review of every diff	If you ask	❌	Varies	✅ Fresh reviewer, P0/P1/P2, optional second vendor
Regression test required for fixes	If you ask	❌	Varies	✅ Always
Agent holds your credentials	Yes, your session	Often	Yes	❌ Never — an orchestrator holds them
Can reach your default branch	Yes	Depends on the token	Varies	❌ Never, at any autonomy level
Stops to ask mid-task	Yes	Fails instead	Varies	❌ Parks the ticket with one question
Where it runs	Your laptop	Your CI	Their cloud	Your CI, your server, or your laptop
Price	Your subscription	Your subscription or API usage	A monthly fee	Free and MIT licensed; your own usage

By hand, in a terminal

The most flexible option and the best for exploratory work. It needs you there: you start each task, answer its questions and review the result yourself. claudrunner uses exactly the same agent, but gives it a queue, a schedule, a reviewer and a set of rules, so the routine work arrives without you.

One CI workflow per event

A workflow that runs an agent when someone comments on an issue is simple and useful. It reacts to one event, though, and each workflow re-invents the rules: which branch, which tests, when to stop. claudrunner is one install with a shared set of rules, and adds the parts a single workflow lacks — claiming so two runs never collide, an independent review, parking instead of failing, and a sweep that finds work nobody reported.

Hosted coding agents

Hosted agents are convenient and need no setup. They run in someone else’s cloud, usually hold your repository and board credentials, and cost a monthly fee per seat. claudrunner runs where you choose, keeps credentials out of the agent entirely, and costs nothing beyond your own model usage.

When claudrunner is not the right tool

One-off, exploratory work. Use the agent by hand.
Windows-only build machines. The orchestrator is POSIX shell and untested on Windows.
Gitea, Forgejo or Bitbucket Data Center. Not supported yet.