Clicking a button on a Jira ticket and getting a pull request a few minutes later is now a real workflow. The catch shows up in the security review: most connectors run the agent on the vendor's servers, holding a token to your repository and your model API key.
For plenty of teams that ends the evaluation. So here is the shape that passes it: the agent runs in CI you already own, and Jira only pulls the trigger.
You can build this yourself. A Jira Automation rule can send a web request to GitHub's repository_dispatch (or a GitLab pipeline trigger), and a workflow runs Claude Code or Codex with the key from your CI secrets, then opens the PR. It works. Where it gets awkward:
- The CI token sits inside an Automation rule that any project admin can open.
- Nothing comes back. Someone pastes the PR link onto the ticket by hand.
- Nothing stops a looping rule from firing fifty runs before anyone notices.
Disclosure: I built Agent Dispatch for Jira at Katabarwa Labs to close those three gaps. Each project is wired once to a repo, base branch and agent. A "Send to" button on the issue fires your GitHub Actions or GitLab CI, and the status and PR link come back to the issue. The CI token is kept in Forge secret storage, and admins get a site-wide daily run cap, a project allowlist and a CSV of every dispatch.
Honest about the trade-off: the app does call out of Jira, to your CI provider and nowhere else, so it does not carry the "Runs on Atlassian" badge. It never holds your repository or your model key.
Agent Dispatch for Jira on the Atlassian Marketplace
Have you tried an AI coding connector on Jira, and what stopped you or your security team?
