Atlassian's MCP server now exposes 220+ tools and sorts them into read, write and destructive tiers. Copy that split

Atlassian signed a bigger deal with OpenAI, and still refused to make it the default model. That is the lesson.

What happened

Atlassian is expanding its partnership with OpenAI, which dates back to 2023. According to VentureBeat, OpenAI's newest models, including GPT-6 Astra and the GPT-5.6 series, flow into Atlassian's Rovo AI, and an OpenAI representative described the deal as "effectively a spend commitment".

But GPT-6 Astra is not Rovo's default. Rovo runs an internal AI gateway that routes across OpenAI and other providers, balancing capability, speed and cost per task.

The same coverage describes Atlassian's rebuilt MCP server:

• More than 15 million calls a day, per Atlassian.
• 220 plus tools across Jira, Confluence, Bitbucket, Loom, Goals and more.
• Up to 25% fewer tokens on comparable Jira and Confluence work in internal testing.
• Read, write and destructive actions separated into risk tiers, so clients can require explicit confirmation on higher risk steps.
• OAuth 2.1 with per user permissions, plus domain and IP allowlists and audit logging.

Atlassian's own post says teams can connect ChatGPT and Codex through MCP, and that agent sessions are anchored to Jira work items that capture requirements, track agent progress and log decisions alongside human comments.

My take

Two habits here are worth stealing for any client build, even small ones.

First, route by task, not by vendor. A company with a spend commitment to one lab still sends work to whichever model fits the job. If your automations are hard coded to one model, put a thin routing layer in front of them now so you can swap without a rebuild.

Second, the three risk tiers. When I connect an agent to a CRM or project tool, I classify every tool the same way: read, write, destroy. Reads run freely. Writes get logged and spot checked. Deletes and bulk changes wait for a person.

The Jira detail is also smart. Give each agent run a ticket that records what it was asked to do and what it did. When a client asks why a record changed, you have an answer in one place.