Open source · Claude Code plugin · MIT
The autonomous AI coding agent that ships your tickets as pull requests
claudrunner takes ready work from your task board, writes the code, runs your tests, reviews its own diff and opens the pull request. On a slower rhythm it hunts your codebase for security holes, silent bugs and scale traps, and files each one with evidence.
/plugin marketplace add samiraklf/claudrunner
/plugin install claudrunner
/claudrunner:initFree, no account No telemetry Your schedule, your runner
Up and running in three steps
Install the plugin
Two commands inside Claude Code. Nothing to host, no account to create.
Run
/claudrunner:initIt reads your project, finds your test commands, connects your board and asks what it cannot infer.
Pull requests arrive
On the schedule you choose, tickets turn into tested, reviewed pull requests for you to merge.
What it automates
Ticket to pull request
Picks up ready tickets, implements them, writes and runs the tests, and opens the pull request. Nobody has to start it.
Jira guideBug and vulnerability scanning
A regular sweep for injection, broken authorization, tenant leaks, N+1 queries and untested money paths. Every finding cites the file and line.
AI security scanAI code review on every change
A fresh reviewer that did not write the code attacks the diff and grades each finding P0, P1 or P2. Serious ones are fixed first.
AI code reviewRegression tests, always
A bug fix is not finished without a test that fails against the old code and passes against the new one.
How a run worksHands-off, safely
It never stops mid-run to ask. Work that is too big or unclear is parked with one clear question, and the run carries on.
Security modelA cute, funny status dashboard
See your agents' ongoing tasks, reviews and what shipped today, while robots type and the reviewer cracks his rope. Local, GitHub Pages or your own server.
Open the demoTwo loops, on your schedule
The fast loop
Often. Every 10 minutes by default
Fetch ready tickets, claim them, implement, test, run the adversarial review, open the pull request and move the tickets to review. Whatever it could not do is parked with a note.
The slow loop
Nightly by default, or weekly, or your own cron
Sweep the codebase for correctness, security, scale and test gaps. Each finding becomes a ticket with the file, the line, the quoted code, the impact and a proposed fix.
Works with the tools you already use
11 task boards
JiraTrelloGitHub IssuesLinearGitLab IssuesAzure BoardsShortcutAsanaClickUpmonday.comNotion
4 code hosts
GitHubGitLabBitbucket CloudAzure Repos
Any stack
Node / TypeScriptPythonPHPGoJava.NETRustanything else
Safe to leave running
- The agent never holds a credential. A small shell orchestrator talks to your board and your git host.
- It can never push to your default branch, merge, or force-push, at any autonomy level.
- Ticket text is untrusted input. A ticket that tries to instruct the agent is skipped as suspicious.
- It only touches tickets it was given. Every ticket id it reports is checked against its own input.
- Runs never overlap, and every run starts from a clean working tree.
- Saying no is allowed. “Not needed” and “better approach” are real outcomes, with evidence.
Guides
Put the routine work on autopilot
Install it in two minutes. Nothing runs until you say so.