A developer on your team is running Claude inside Cursor. She types "create a Jira issue for the auth regression, link it to the Q4 milestone, and update the Confluence runbook." The MCP server handles it - no tab-switching, no copy-paste. That workflow became meaningfully more capable on October 7 in Amsterdam, when Atlassian announced the Rovo MCP server's expansion to over 220 tools at Team '26 Europe. The detail buried in the changelog: every one of those calls spends Rovo credits.
What the Atlassian MCP server actually does now
Atlassian MCP now connects more Atlassian products to the AI clients you choose, and the server was redesigned from the ground up for efficiency - delivering 220+ tools without bloating context.
A single server gives AWS, ChatGPT, Claude, Cursor, Devin, Docker, Figma, GitHub, Google, Lovable, Mistral, Postman, VS Code, and others access to Jira and Confluence.
The jump in tool count matters more than the headline suggests. A test organization after the October announcement showed 36 toolsets listed in the MCP Permissions tab, 19 tagged New - every read, write, and search toolset allowed by default, with delete and manage blocked.
The suite covers Jira, Jira Service Management, Confluence, Bitbucket, Projects, Goals, and Loom.
The practical Jira tools are concrete:
searchJiraIssuesUsingJql, getJiraIssue, createJiraIssue, editJiraIssue, transitionJiraIssue, and manageJiraSprint on the Jira side; Confluence adds page reads, writes, comments, moves, and archives.
Atlassian's internal benchmarking on Claude models found the redesigned server uses up to 25% fewer tokens per call
- relevant if you're hitting context limits on long sprint summaries or multi-issue searches.
The auth story is cleaner than it was in beta:
the server runs at mcp.atlassian.com/v2/mcp with OAuth 2.1, and API tokens are also accepted.
Delete operations - deleteJiraIssue, deleteJiraComment, and attachment deletes - are disabled by default, and the server respects each user's existing Jira and Confluence permissions.
That last part is the one governance detail that matters most: an agent can only see and write what the authenticated human could.
The Rovo credit mechanic nobody is budgeting for
This is the part most coverage skips. Tool calls through the Rovo MCP server consume Rovo credits. Atlassian says most tool calls cost 1 to 10 credits, and heavier generation tasks use 15 to 30.
Think through what that means in practice. A developer who asks Claude to "find all open P1 bugs, create a tracking issue, and update the status page" might chain four or five tool calls - a search, a read, a create, a write - each spending credits. At 10 credits per call on the high end of routine operations, that's 40-50 credits for one natural-language request that used to take three minutes of tab-switching. If your team runs twenty requests like that a day, you're burning roughly 1,000 credits daily before any generation tasks.
Atlassian hasn't published a per-credit dollar rate in the public docs, so you can't easily convert this to a monthly line item without checking your specific plan. What you can do right now: go to Atlassian Administration → Rovo → Rovo MCP server → Permissions and audit which toolsets are enabled. Eight new write toolsets were on by default in the October rollout:
write_goals, write_projects, and write_teams were among them.
Turn off any you don't need. A write toolset your team isn't using is just credit exposure with no upside.
Beagle in action#engineering, 10:47amThe v1 deprecation clock
The v1 endpoint deprecates on March 1, 2027.
Five months sounds comfortable. It isn't, if you have automation scripts, CI pipelines, or internal tools hardcoded to the old URL. The v2 endpoint is mcp.atlassian.com/v2/mcp. If your MCP client config reads mcp.atlassian.com/mcp anywhere, that's a migration item, not a backlog item.
Atlassian also added new capacity planning tools in late September - read_capacity_planning and write_capacity_planning toolsets
- so teams doing sprint planning through an AI client now have direct read/write access to individual allocation data. Worth knowing before an agent starts reshuffling sprint capacity on an engineer's behalf.
What teams should configure before going further
The Atlassian MCP server is genuinely useful as-is. It's also the kind of tool that benefits from fifteen minutes of setup discipline before you let it loose.
| Step | What to do | Why |
|---|---|---|
| Audit write toolsets | Disable any write_* toolset your team won't use |
Limits credit spend and unintended writes |
| Check auth scope | Confirm each user connects with their own OAuth, not a shared service account | Preserves permission boundaries |
| Update MCP client config | Point to mcp.atlassian.com/v2/mcp |
Avoids March 2027 breakage |
| Monitor credit burn | Check Rovo credit usage weekly for the first month | Surface unexpected call volume early |
| Pin delete toolsets to off | Verify deleteJiraIssue and similar are blocked |
Default is off, but verify after any toolset update |
The capacity planning toolsets deserve a separate call: if you're enabling write_capacity_planning, decide who gets that scope. An agent that can write allocations across projects is touching something that usually has a dedicated person behind it.
A teammate like Beagle can help surface when a new write toolset appears and draft the internal review thread - but the approval decision needs a human with Atlassian admin access and a read of the changelog.
Atlassian MCP server: common questions
What AI clients work with the Atlassian MCP server?
The server supports AWS, ChatGPT, Claude, Cursor, Devin, Docker, Figma, GitHub, Google, Lovable, Mistral, Postman, VS Code, and others
- any client that speaks MCP. Auth is OAuth 2.1 or an API token; you connect once and the server enforces your existing Jira and Confluence permissions.
Does the Atlassian MCP server cost extra beyond my Atlassian plan?
Tool calls through the Rovo MCP server consume Rovo credits. Most tool calls cost 1 to 10 credits; heavier generation tasks use 15 to 30. Whether those credits are included in your plan or billed separately depends on your Atlassian subscription tier. Check Atlassian Administration → Rovo for your credit balance and usage history.
What is the difference between Atlassian MCP v1 and v2?
Atlassian Rovo MCP v2 became generally available on September 8, 2026, and introduced significant changes including the expanded toolset, a redesigned token-efficient architecture, and OAuth 2.1 as the primary auth method. The v1 endpoint deprecates on March 1, 2027.
Can an AI agent delete Jira issues through MCP?
Delete operations - including deleteJiraIssue, deleteJiraComment, and attachment deletes - are disabled by default, and the server respects each user's existing permissions.
You can enable them explicitly in the Permissions tab, but there's no good reason to for most teams.
Which Atlassian products does the MCP server cover?
The suite covers Jira, Jira Service Management, Confluence, Bitbucket, Projects, Goals, and Loom
- all from a single endpoint at
mcp.atlassian.com/v2/mcp. A single OAuth authorization grants access to whichever products your account can reach.