Atlassian OpenAI Partnership Expands Rovo and Codex

The Atlassian OpenAI partnership expanded on October 6, 2026, through an agreement bringing OpenAI frontier models to Atlassian agents and Rovo. Existing Atlassian and Teamwork Graph plugins connect project context to ChatGPT and Codex, subject to permissions. OpenAI announcement.

Atlassian OpenAI partnership: what changed

The agreement expands Atlassian’s access to models including GPT-6 Astra and the GPT-5.6 series. Rovo combines model reasoning with Teamwork Graph, which connects people, projects, documents and decisions. OpenAI reports that more than 3,000 Atlassian developers use Codex. Partnership details.

Our assessment: teams should evaluate the collaboration through a specific decision they need to make. For a release manager, that might be identifying which unresolved issue changes a launch date. For an engineer, it might be checking whether the requested change matches the approved specification. Define that question before choosing a model or connecting a larger collection of records. Our GPT-6 Astra overview provides separate model background.

Choose a supported surface for existing plugins

OpenAI’s current plugin documentation lists ChatGPT Chat and Work on web, desktop and mobile, Codex in the ChatGPT desktop app, and the Codex CLI plugin browser. The IDE extension does not support plugins. Plugins marked Desktop only require the desktop app.

On web or desktop, inspect the plugin in the Plugins tab, install it and complete any requested service connection. Start a new chat to use its capabilities. In Codex CLI, open the plugin browser with /plugins, install from a configured marketplace and begin a new session. Plugin details and connection prompts determine the setup you actually need.

For a team pilot, we recommend recording the intended client and task together: “Codex CLI, compare the release issue with its linked specification,” for example. Test that pairing as the member who will perform the work. A demonstration on another client does not answer whether the intended workflow is usable.

Match project context to the authorized account

OpenAI’s plugin controls guide separates plugin availability, MCP server access, permitted actions, service authorization and runtime permissions. Members need access to the plugin and server through their role, plus access to the external service through the authenticated account. Shared credentials require permission to use that connection.

Installing a plugin across a workspace does not supply a shared account or grant access to every external record. Where supported, administrators can configure read actions or a custom action set and control when approval is requested. Codex host sandbox and approval policies also apply to plugin capabilities, as explained in the plugin guide. The expanded partnership should therefore be evaluated against the actual account and workflow.

Our recommendation is to write down the intended project, account and allowed action before the first request. For the launch review below, the required action is reading selected evidence and producing a draft assessment. Keep any subsequent issue edit as its own decision. Our Enterprise plugin administration report covers the earlier console update.

Untested editorial example: review one launch decision

The following fictitious exercise is our proposed evaluation, not a reported Atlassian deployment or a tested result. Imagine a browser account switcher scheduled to launch on Friday. Select three authorized Jira issue records and two short Confluence decision documents. Supply them through an approved connection or as selected excerpts, depending on the workflow you are evaluating.

Use this small fixture:

• SW-21: interface merged; the linked acceptance checklist still needs confirmation.
• SW-22: permission review open, assigned to Maya, due Thursday.
• SW-23: regression test failed; no owner or retry date recorded.
• Launch brief: Friday is the target, conditional on permission approval and a passing regression test.
• Decision note: migration of existing login sessions belongs to a later release.

Ask for a launch assessment with one row per release condition. Each row should show the evidence identifier, current state, missing information, and the next decision required. Ask the model to preserve Friday as a target and distinguish recorded owners from proposed assignments. Request a separate list of questions for the release manager, without editing issues or sending messages.

Inspect three things in the answer. First, “interface merged” should not become “all interface checks passed.” Second, SW-22 should retain Maya and Thursday, while SW-23 should keep its missing owner and date visible. Third, the deferred migration should remain outside the current release. A fluent summary that erases any of those distinctions would not satisfy this exercise.

Now change just one input: replace the failed regression record with a passing result that includes its test identifier and timestamp. Repeat the same request. The regression condition should change, while the pending permission review should continue to affect the launch decision. This tests whether the assessment follows the supplied evidence rather than repeating its first conclusion.

Our guide to source-linked decisions and actions offers a reusable method for preparing and checking these inputs. Save both outputs with the fixtures so the team can compare the exact reasoning presented to the reviewer.

Measure the workflow before expanding it

We recommend judging this pilot by decision quality: were all release conditions retained, did every claimed status point to evidence, and did the reviewer receive the unresolved questions needed to decide? Record the time spent preparing inputs, checking the answer and correcting errors. Compare it with the same task performed through the team’s existing process before claiming a benefit.

Keep the sample consistent when trying another client or model. Otherwise, a cleaner input or easier project could explain the difference. Expand to another release only after the team can identify what the first exercise established and what still needs checking.

Deeper Jira integrations for assigning work to agents, tracking progress and reviewing results remain under exploration. The developer adoption figure reports usage, not measured productivity gains. OpenAI’s roadmap discussion.