Atlassian and OpenAI: MCP server, 220 tools, Rovo stays multi-model

Atlassian and OpenAI expanded their partnership on October 6, and the headline is the money: Atlassian is committing spend to OpenAI models for Rovo and its agents. The detail we care about is smaller. Atlassian rebuilt its MCP server, it now exposes more than 220 tools across Jira and Confluence, and Rovo still routes work across several model vendors instead of betting on one.

What actually happened

Per OpenAI's announcement and VentureBeat's reporting:

• Models. Atlassian gets expanded access to OpenAI frontier models, including GPT-6 Astra and the GPT-5.6 series, to power agents across its platform and Rovo.
• Money. An OpenAI representative told VentureBeat the deal is "effectively a spend commitment" from Atlassian. Contract terms were not disclosed.
• MCP server. The rebuilt Atlassian MCP server handles more than 15 million calls a day and exposes 220+ tools across Jira, Confluence, Bitbucket, Loom, and Goals. Atlassian says it used about 25% fewer tokens than the old version in internal tests. Jira Service Management support is listed as coming soon.
• Connectors. OpenAI says the MCP connectors move data between ChatGPT, Codex, and Atlassian tools.
• Not exclusive. VentureBeat reports that Rovo runs an internal AI gateway that picks a model per task, GPT-6 Astra is not the default, and Anthropic's Claude models are still in use.
• Controls. The server is Atlassian-hosted with OAuth 2.1, enforces user permissions at query time, and splits actions into read, write, and destructive tiers, with audit logging.

Why it matters for your business

If your team lives in Jira and Confluence, the plumbing for "ask an agent to update the ticket" is now the vendor's job, not yours. Two things follow.

1. MCP is the integration layer now. 220 tools behind one standard protocol means ChatGPT, Codex, Claude, or your own agent can reach the same Jira data. You pick the client. You do not rebuild the connector each time.
2. Copy the gateway, not the deal. A company with Atlassian's leverage signed a big OpenAI commitment and still routes by task. Price makes the case: VentureBeat lists Astra at $10 per million input tokens and $50 per million output. Most ticket triage does not need that. A router that sends easy work to a cheap model and hard work to a frontier model is the cost control.
3. Use the permission tiers. Read, write, and destructive are separate. Start agents on read. Grant write per project. Keep destructive actions behind a human.

Sources: OpenAI, VentureBeat.