OpenAI announced an expanded Atlassian partnership on October 6, 2026, naming GPT-6 Astra and the GPT-5.6 series among the models Atlassian can access. The companies say OpenAI models will power agents across Atlassian’s platform and Rovo. For product, engineering and IT teams using Atlassian Cloud, the practical question is how to connect Jira and Confluence work to ChatGPT, Codex or an Atlassian-managed Rovo agent.
The agreement is new; not every integration is. Jira’s coding-tool deeplinks launched on June 16, 2026. OpenAI describes deeper Jira integration for assigning and tracking agent work, alongside DX-based impact measurement, as exploratory work, not a committed release.
Choose the work environment, then the connection
Atlassian’s Teamwork Graph connects people, projects, documents and decisions into organizational context, according to OpenAI’s announcement. A useful way to evaluate the options is to separate where the user works from how the agent gets that context.
MCP and Teamwork Graph CLI provide ongoing access to Jira and Confluence context; a Jira deeplink starts a task. That distinction follows from Atlassian’s CLI/MCP comparison and deeplink documentation. MCP means Model Context Protocol: an interface through which an AI client can use external tools.
Your immediate task | Start with | Boundary to understand |
|---|---|---|
Summarize project work or prepare a status answer | Select relevant work in the prompt and connect the permitted Atlassian identity; installation alone does not establish access. | |
Plan or implement a repository change using work requirements | Retrieve Jira and Confluence context through supported tools. Repository access and permission to update Atlassian content are separate. | |
Chain commands or process graph context in a shell-capable workflow | Requires a usable local binary and cloud connectivity. Check command coverage; it is not an offline deployment. | |
Start coding from one well-specified Jira item | The documented handoff is the summary and description. Optional MCP adds further retrieval and updates; the link alone is not ongoing synchronization. | |
Configure an agent within Atlassian’s managed experience | Specific model choices belong to Think deeper reasoning. Inspect the options actually offered to your tenant. |
For Codex, choose one primary access mechanism per session. Atlassian recommends making the CLI/MCP choice explicit when both are installed. Use MCP when the host supports it but cannot reliably run a shell; use the CLI when scriptable outputs and local tool chaining are part of the task.
Confirm availability in the actual workspace before committing to a route. OpenAI’s current plugin documentation distinguishes supported ChatGPT and Codex clients from the IDE extension, which does not support plugins. The IDE can still connect directly to MCP. For Rovo, plan, region and configuration can affect model options; the partnership does not establish universal GPT-6 availability.
A small starting point for Codex
For an approved Codex MCP rollout, Atlassian documents the v2 endpoint in its setup guide. The login and inspection commands below come from OpenAI’s MCP documentation:
codex mcp add atlassian --url https://mcp.atlassian.com/v2/mcp
codex mcp login atlassian
codex mcp listComplete the sign-in flow, inspect the available tools, then retrieve one known work item and its linked decision document before asking for an implementation plan. Do not enable write access merely to prove retrieval works.
Constrain actions separately from data access
A user may be entitled to read dozens of projects while a release-readiness task needs only one. Our rollout recommendation is to limit the task’s data scope and executable actions separately. A prompt asking an agent to stay read-only is not a substitute for disabling unnecessary write capabilities.
For MCP: Atlassian admins can configure Read, Write and Search permissions, including per-app settings. Review the option to apply permissions to future additions as well as today’s permissions; a broad grant can cover capabilities added later.
For TWG CLI: Read, Write and manage, and Delete are separate categories. The permission controls also distinguish saving a restriction from revoking existing sessions. Atlassian recommends session revocation when removing permissions.
For the AI client: Use the host’s tool restrictions and approval settings alongside Atlassian’s controls. Codex supports tool allow/deny lists and approval policies. Before permitting a mutation, review its exact target and proposed change through the team’s normal process.
Check who receives the content
Atlassian’s partnership guide distinguishes content sent to ChatGPT or Codex, where OpenAI-side processing follows the organization’s OpenAI plan and terms, from models used within Atlassian-managed Rovo. Choosing Rovo does not mean every model runs on Atlassian infrastructure: its data guidelines describe third-party hosted models too.
Atlassian says those model providers do not retain inputs and outputs or use them to improve their services; Atlassian itself retains Rovo Chat and agent inputs and outputs for 30 days for safety and security. Those are Atlassian’s service-policy statements, not guarantees that automatically transfer to a separate ChatGPT or Codex connection.
The implementation question is which fields leave the source system, who receives them and for what purpose. RohitAI’s guide to agent data-flow review explains that distinction. Installing a CLI locally does not make the external model processing local.
Budget for context as well as the model
Atlassian’s current Rovo rate card lists single-product lookups and writes as free context operations. Enriched calls through MCP or TWG CLI consume Rovo credits; most cost 1–10 credits, not a guaranteed maximum. Allowances are pooled across the organization. Extra-usage billing is scheduled for December 3, 2026, at US$0.01 per credit. Admins can disable extra usage or set a spending cap.
Illustrative calculation, not a measured bill: assume 100 tasks each make four enriched calls, all within that 1–10-credit range. That is 400 calls × 1–10 credits = 400–4,000 credits. If every credit exceeds the included allowance after extra-usage billing begins, the context overage would be US$4–$40 at the published rate. This excludes OpenAI usage, subscriptions, discounts, retries and other Rovo activity. Calls outside that common range can change the result; it is neither a ceiling nor a forecast.
For a purchasing decision, compare total spend per accepted result, not model-token savings alone.
Start with one reviewable deliverable
A useful first rollout is a cited status answer in ChatGPT, a repository implementation plan in Codex, or one narrowly scoped Rovo task. Our suggested acceptance criteria are:
The answer links to the Jira items and Confluence pages supporting it, and identifies missing or inaccessible evidence.
The effective identity, permitted data and enabled actions match the intended task; no unexpected updates occur.
The team records context usage and model usage separately before widening access or enabling changes.
A successful sign-in is only the connection check. Advance when the output is useful, its evidence is inspectable and its permitted actions match the workflow.
Methodology: AI-assisted reporting and analysis based on official OpenAI and Atlassian sources rechecked on October 6, 2026. No hands-on testing or independent productivity measurement was performed. Recommendations and calculations above are analytical, not observed outcomes.