Nine out of ten developer teams now use AI for coding. Overall productivity gains have plateaued at 10 to 15 percent. That's the exact gap Claude Agent for Jira is designed to close.
Launched on June 18 in partnership with Anthropic, the integration lets you assign a Jira work item directly to Claude. A few minutes later, a draft pull request is waiting. Not an autocomplete suggestion, not a chat response. An actual code change, on its own branch, ready for your team to review.
What does Claude Agent for Jira actually do?
You assign a Jira ticket to Claude the same way you'd assign it to a teammate: tag Claude in a comment, pick the agent from the assignee menu, or set up an automation rule to route tickets automatically. The rest happens in the background.
You need a Jira Cloud subscription (Standard, Premium, or Enterprise) with Rovo enabled. That's the only prerequisite before you can try it via the Atlassian Marketplace.
How does a ticket become a pull request?
Once you assign the ticket, Claude starts a sandboxed session in Anthropic's isolated infrastructure, clones your repository, and reads the codebase before changing a single line. It then implements the changes on a separate branch and opens a draft pull request.
For context: you link the GitHub repository upfront. Claude analyzes your existing code and style first, so the pull request looks like something your team would write, not an unfamiliar chunk dropped in from outside. The sandboxed environment means Claude has no access to your production systems or secrets.
Think of it like handing a task to a new contractor. They work on their own copy, deliver it as a proposal, and you approve only after review. Live progress appears on the ticket card as Claude works, so you can see the agent think through the problem in real time.
Critical detail: Claude stops at draft. It merges nothing on its own. Human review remains the final step, and that's deliberate.
Why is this more than another AI button?
The difference isn't about who has a hand in writing code. It's about who finishes the work.
A longitudinal study across 400 engineering organizations found that AI adoption for coding tasks reached 90%, but productivity gains plateaued at 10 to 15%. The pattern is familiar in most teams: everyone has a chat window open, but the actual work, the ticket, the branch, the review, still runs exactly as it did before. The gains leak away in context switching.
By putting the agent inside Jira, where work already lives and is tracked, Atlassian is trying to close that gap. Whether it works depends less on the technology and more on the discipline around it. Because AI writes a lot of code, and not all of it holds up.
Anthropic's own data shows that roughly 34% of Claude code sessions by software engineers complete fully and are verified as correct. That's exactly why the draft pull request is the right design choice: the output isn't the code itself, it's the proposal a human still evaluates.
What does it cost?
You pay in two places: your Jira and Rovo subscription, and the compute for the agent's work. Core Rovo is included in paid Jira Cloud plans, with 25, 70, or 150 Rovo credits per user per month on Standard, Premium, and Enterprise respectively.
For heavier developer work, there's Rovo Dev at $20 per developer per month, which includes 2,000 separate credits monthly. Beyond that, Atlassian charges $0.01 per extra credit. If you use the local route through your own coding tool, you pay model usage through your own Claude subscription or API key.
For a five-person developer team on Rovo Dev, that's $100 per month before any overage. The harder number to predict is how many tickets you actually route through the agent. That second layer is where costs can climb quickly, so run the numbers before enabling it for the whole team.
Open tickets directly in Claude Code or Cursor
Separate from the Jira agent, a June 16 update lets you open any work item in your preferred coding tool with one click. In the ticket's Development panel, select "Open in coding tool" and choose from Claude Code, Cursor, GitHub Copilot, OpenAI Codex, VS Code, or the Rovo Dev CLI.
The feature runs on the Atlassian MCP, the Model Context Protocol that passes context between tools. The ticket summary and description transfer automatically. Desktop apps like Cursor open pre-loaded with the context; terminal agents like Claude Code receive a window with the complete prompt ready to copy.
It removes the copy-paste ritual that starts every ticket-to-code session. That sounds minor. It's the kind of friction that adds up across an entire workday.
How does this compare to Cursor and Copilot?
Claude is not the first coding agent to land in Jira, but it is the first that takes over a complete ticket from within Jira itself. The key distinction runs between two working modes: the cloud agent that runs in Atlassian's infrastructure and delivers a pull request, versus the local agent you direct on your own machine.
| Factor | Claude Agent for Jira | Local agent via deep-link |
|---|---|---|
| Where work happens | In Atlassian's sandbox | On your own machine |
| How you start | Assign ticket to Claude | Click "Open in coding tool" |
| What you get back | Draft pull request | Tool opens with context loaded |
| Who pays for compute | Your Rovo plan | Your own Claude or AI subscription |
| Best for | Routine, well-specified tickets | Complex work you want to steer |
The two approaches don't exclude each other. You can hand a simple, well-defined ticket to the cloud agent and open a more complex card in Cursor at the same time. Same control panel, two speeds.
When should you not hand Claude your ticket?
Claude delivers a draft, not a finished merge, and for some tickets that draft won't be worth much.
A vague ticket produces a vague pull request. Claude reads the summary, description, and acceptance criteria. Without a sharp spec, it guesses. And it guesses confidently.
Work that requires an architecture decision or touches security shouldn't go to an unattended agent. The sandbox won't reach your production systems, but a plausible-looking pull request built on the wrong assumption still costs you review time. Rigorous pull request review before merging remains essential.
Also avoid tickets that rely on knowledge that exists nowhere in writing. Whatever isn't in Jira or connected Confluence, Claude can't access, and it fills those gaps with assumptions. That's exactly where a polished-looking solution quietly misses the point.
The teams that get the most from this are the ones who already write clean, complete tickets. If your backlog is full of one-liners, you're not solving the problem, just pushing it to the review stage.
What you can do on Monday
Pick one small, well-defined ticket from your backlog and assign it to Claude to see what comes back. Choose something low-stakes intentionally: a bug fix, a cleanup task, a routine change with clear acceptance criteria.
Check that Rovo is enabled on your Jira Cloud subscription, then treat the draft pull request like any junior's first submission: read it, test it, ask questions, then merge. Claude appeared earlier as a teammate in Slack, and the pattern is the same: the AI does the prep work, and you hold the final call.
The full announcement is at Atlassian and in the Jira launch notes. More on Anthropic's agent approach is at anthropic.com.
