Integrations
Run Claude Code on Jira tickets with a self-hosted worker
This guide pairs Jira Cloud with Claude Code. DevIntern is the orchestrator in between: a self-hosted worker that polls your JQL, runs Claude Code on each ready issue on your machine, and opens a pull request. Below is one ticket traced end to end, the exact config for this pairing, and the Claude Code settings that matter for unattended runs.
Looking for the general setup? See the Jira integration overview or the Claude Code harness overview .
Claude Code is fast. Feeding it Jira tickets is not.
Most teams already pay for Claude Code, yet every ticket still needs someone to copy the description, attach screenshots, create a branch, run the agent, and open the pull request. That hand-off is the slow part, and it is the part DevIntern runs for you as a background agent.
Jira stays the source of truth. Your JQL decides what is ready, your workflow statuses show progress, and the outcome of every run is written back to the issue as a comment.
One Jira ticket, start to finish
Here is what happens to PROJ-123 once someone adds the devintern label in the open sprint, using the config further down this page.
- Jira issue
- Worker on your machine
- Claude Code
- Pull request on your repo host
- Review comment
- New commit
-
1. The worker notices the change
Every poll_interval (60 seconds by default) a cheap change detector asks Jira whether anything changed. Only then does the worker re-run your JQL. The issue matches, and the worker log shows the pickup.
[poll:jira] 📌 picking up PROJ-123 -
2. Claude Code checks the ticket before writing code
DevIntern converts the ADF description to markdown, includes comments from teammates (its own earlier comments are filtered out), and downloads attachments. The feasibility check runs in Claude Code's read-only plan mode, capped at 10 turns. A vague ticket gets an "Automated Task Feasibility Assessment" comment with concrete questions and no pull request.
-
3. Implementation in a disposable worktree
The issue moves to your inProgressStatus. Claude Code runs headless in a fresh worktree from a managed clone, up to --max-turns (default 500). Your own checkouts are never touched.
-
4. Pull request and Jira update
DevIntern commits, pushes the branch, and opens the pull request. The issue moves to prStatus and gets an "Implementation Completed by @devintern/code" comment. The worker records the issue as actioned, so later polls skip it even if it still matches your JQL.
[poll:jira] ⏭️ skipping PROJ-123 (already actioned; no change since the PR) -
5. Review comments become commits
When a reviewer requests changes or leaves inline comments on the pull request, the worker runs Claude Code on the feedback and pushes new commits. Review feedback is handled before new tickets are picked up. With [workspace].ci_failure_fix = true it also asks Claude Code to repair failing CI on its own pull requests.
-
6. If the run cannot finish
The issue gets an "Implementation Incomplete" comment and moves back to todoStatus. The worker will not retry an unchanged ticket in a loop: edit the description, post a comment, or delete the failure comment to unlock a retry. When the previous attempt left a branch and report, the retry picks up where it stopped.
The full config for Jira and Claude Code
Install DevIntern and create the worker with the quick start, then use these three files. Credentials stay in a local .env that is kept out of version control. Read the quick start.
Workspace .env
TASK_TRACKER=jira
JIRA_BASE_URL=https://your-org.atlassian.net
JIRA_EMAIL=you@example.com
JIRA_API_TOKEN=your-api-token
AGENT_HARNESS=claude-code
# Optional: a Claude Code model alias and reasoning effort
AGENT_MODEL=sonnet
AGENT_EFFORT=medium
# Optional: Claude Code's own sandbox, configured by DevIntern
AGENT_SANDBOX=native JIRA_BASE_URL is your Jira Cloud site without a trailing slash. JIRA_EMAIL must be the Atlassian account that owns the API token. Jira Server and Data Center are not supported.
.devintern-code/settings.json
{
"jira": {
"projects": {
"PROJ": {
"inProgressStatus": "In Progress",
"todoStatus": "To Do",
"prStatus": "In Review"
}
}
}
} The key is the project key from issue keys (PROJ in PROJ-123). Status names must match your workflow exactly. Omit a field to skip that transition.
workspace.toml
[defaults]
tracker = "jira"
task_query = "sprint in openSprints() AND labels = devintern AND status = 'To Do'"
worker_task_args = "--create-pr"
poll_interval = 60
# Optional: only pick up new tickets overnight
[worker.schedule]
active = ["22:00-06:00"] Before turning the worker on, try one issue by hand: devintern PROJ-123 --create-pr. Quiet hours pause new pickups only; review replies keep running.
Claude Code settings for unattended runs
DevIntern finds the claude binary on your PATH and uses the login or API key it already has. These are the knobs worth knowing before you leave it running.
Model and effort
AGENT_MODEL takes a Claude Code model alias such as sonnet and applies to every spawn. AGENT_EFFORT accepts low, medium, or high; lower is faster and cheaper, higher reasons more deeply.
Turn limits
Implementation runs default to 500 turns; raise it for large refactors. Feasibility checks always use 10.
worker_task_args = "--create-pr --max-turns 1000"Native sandbox
With AGENT_SANDBOX=native, DevIntern generates a Claude Code settings file per run: the Bash-tool sandbox is on, unsandboxed retries are off, writes are limited to the worktree, task output, and temp directories, and docker and gh are excluded because they break under the sandbox. It confines shell commands, not Claude Code's built-in file tools; use nono or srt for a boundary around the whole process. Run devintern sandbox to see what your next run will do.
Failover when you hit a usage limit
List a second harness and the worker switches when Claude Code hits its limit. Failover covers tickets, review replies, and conflict resolution.
AGENT_HARNESS=claude-code,codexCommon problems with this pairing
"Claude CLI not found"- Install Claude Code and make sure claude is on the worker's PATH, or set AGENT_CLI_PATH. devintern doctor checks this and suggests an installed alternative.
Jira API error (401)- The API token is invalid or revoked, or JIRA_EMAIL is not the account that owns it. Create a new token in your Atlassian account settings.
The issue matches the JQL but is not picked up- Check the [poll:jira] lines in the worker log. "already processed at this update" means edit or comment on the issue; "already actioned" means a PR exists, so edit the summary, reopen it, or change a label to re-arm it.
Status transition failed- The status name in settings.json does not match your workflow, or the transition is not valid from the current status. DevIntern logs a warning and keeps going.
Sandboxed run fails to start on Linux- Claude Code's native sandbox needs bubblewrap on Linux, and DevIntern makes the run fail instead of running unsandboxed. Install it, or pick another provider after checking devintern sandbox.
Hosted alternatives
Atlassian offers Rovo Dev in Jira, which runs in Atlassian's cloud on Atlassian-selected models with Rovo Dev credits. It is a good fit if you want a fully managed option inside Jira.
DevIntern fits teams that want Claude Code on their own subscription, running on machines they control.
Source: https://www.atlassian.com/blog/company-news/rovo-dev-in-jira
FAQ
- Can I use my Claude subscription?
- Yes. DevIntern drives the Claude Code CLI you already installed and signed in, so runs use your existing subscription or API key. It does not host or proxy model calls.
- Is DevIntern a Jira coding agent?
- No. Claude Code is the agent. DevIntern is the self-hosted ticket-to-PR worker around it: polling Jira, managing worktrees and branches, opening pull requests, and writing results back to the issue.
- Can I keep it from commenting on issues?
- Yes. --skip-comments skips all Jira comments and status transitions for a run while still implementing the change.
Try it from your terminal
Interactive use is free. Run one ticket by hand, then turn on the worker when you want it to pick up tickets unattended.
Related: Jira integration overview , Claude Code harness overview , Worker docs , Linear and Codex , Self-hosted vs hosted coding agents