Quick answers
- What is the Atlassian MCP server?
- A cloud-hosted MCP server run by Atlassian that connects "your AI agent to your Atlassian apps and work": Jira, Jira Service Management, Confluence, Bitbucket and Loom.
- Whose permissions does the AI assistant use?
- With the standard sign-in (OAuth 2.1), the person who signed in, narrowed by the scopes they approved and your admin's MCP settings. With an API token, the token's owner.
- Which Atlassian deployments does it work with?
- Atlassian Cloud: the prerequisites list "An Atlassian cloud site with Jira, Compass, and/or Confluence".
What the Atlassian MCP server is, and what it replaced
MCP (Model Context Protocol) lets an AI client call tools on another system. Atlassian's tools cover JQL search, Confluence content, editing and transitioning work items, and Teamwork Graph.
On 7 October 2026 Atlassian said "Atlassian Rovo MCP is now Atlassian MCP". Version 2 lives at https://mcp.atlassian.com/v2/mcp. From 1 March 2027, v1 connections move to v2 tools automatically, and you'll need to authenticate again.
MCP calls can draw on your Rovo credits; single-product lookups and updates are free. Our Rovo and Atlassian AI page covers credits.
How to connect Claude, ChatGPT and other clients
As documented on 19 October 2026. With the standard sign-in, each route ends at an Atlassian consent screen.
| Client | How to connect | Who acts first |
|---|---|---|
| Claude (web and desktop) | Browse connectors, select Atlassian, Connect, sign in. Not from iOS or Android. | Your Atlassian admin, if AI domains are restricted |
| Claude Code | claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp, then /mcp | No one by default; project scope shares it via .mcp.json |
| ChatGPT | OpenAI lists "Atlassian Rovo (Legacy)"; v2 needs Atlassian's new plugin. Check which. | Your ChatGPT workspace admin |
| Other MCP clients | Point Gemini, Copilot CLI, VS Code, Cursor or Codex at the v2 endpoint | Your Atlassian admin, if the domain isn't allowed |
How to verify your connection
Use a test project with nothing sensitive in it.
- 01
Read something
Fetch a test work item by its key. No data? Re-authenticate or check the granted scopes.
- 02
Make one harmless change
Comment on a test work item, check which account Jira shows as author, then delete it.
- 03
Check what it can reach
Compare the tools your client lists with what you expected, and ask your admin to confirm MCP permissions.
- 04
Find it in the audit log
In Atlassian Administration, Insights, Audit log, search "MCP": each entry shows the tool, action and user.
If it doesn't work
- "You don't have permission to connect from this IP address."Your IP allowlist is blocking the AI tool's network.
- "Access denied", naming a permissionAn admin has blocked Read, Write or Search for the MCP server.
- "You don't have permission to connect via" a methodYour admin hasn't allowed that sign-in method, such as API tokens.
Permissions and data flow: what to understand before you switch it on
With OAuth 2.1, the assistant is you, within limits: your access, narrowed by approved scopes, the tools on offer, your admin's MCP permissions and, for reading, any data security policy. With an API token, it's the token's owner, and audit entries may not show each person.
Content returned to the AI client is then governed by your agreement with that vendor; Atlassian's MCP pages don't describe retention there.
What your Atlassian admin controls
On Government Cloud or Isolated Cloud, confirm MCP availability with Atlassian first.
- MCP permissionsAllow or block Read, Write and Search per app, under Connectors; changes apply immediately.
- Data security policyWith Guard Standard, block MCP reading of Jira and Confluence; OAuth only, not API tokens.
- DomainsAllow or remove AI tool domains; this applies only to OAuth 2.1 tools.
- API token sign-inOn or off; Jira Service Management operations tools currently need API tokens.
- High-impact toolsdelete_jira and manage_jira are "disabled by default" until an admin enables them.
- Confluence sharing toolsPermission and public-link tools share Confluence write, so only blocking write stops them.
- IP allowlistsYour existing allowlists still control where users and tools connect from.
- Audit logEvery tool call is logged on all tiers; logging the first OAuth authorisation needs Guard Standard.
Which route fits: Atlassian MCP, Rovo or another MCP server
| Route | Start here when | What decides it |
|---|---|---|
| Atlassian MCP server | Your people already work in an MCP client | Your AI vendor's terms and your permissions |
| Rovo, inside Atlassian | You want AI where the work happens | Plan and data handling: see our Rovo page |
| Teamwork Graph CLI | You're scripting or running CI/CD | Whether a person is in the loop |
| Community or self-hosted servers | You need what Atlassian doesn't expose | Who maintains it, its credential, vetting each release |
What the MCP does well, and what it can't decide
Claude is quick at reading Jira and Confluence and standing up simple structure. Building a BAU project on our own Jira, the board stood up quickly. The automation rules Claude produced created a loop where a task couldn't be closed, and kept reaching for labels, which are hard to scale.
We've seen a similar split in customer teams: simple design goes well, permissions and automations often fall apart. Atlassian's tool list has no tool for automation rules or custom fields, and the MCP can't tell you which design suits your environment, your reporting, or your team in two years.
AI speeds up building what's been decided; deciding is architectural judgement. Our Atlassian consulting team works through it with you, and our managed services can check the build matches it.
Before you roll it out: an admin checklist
- Allow only the AI clients and domains you've chosen.
- Set Read, Write and Search per app; check auto-granting of new permissions.
- If API token sign-in stays on, name each service account, its reach and owner.
- Leave delete_jira and manage_jira off without a reason and a reviewer.
- Review Confluence space and Jira project permissions before anyone connects.
- Confirm the networks your AI clients connect from against your IP allowlists.
Frequently asked questions
Does the Atlassian MCP server work with Jira Data Center?
Atlassian documents it for Atlassian Cloud sites. The setup pages we checked on 19 October 2026 don't cover Data Center.
Can Claude or ChatGPT delete Jira issues through it?
Not by default: delete_jira "is disabled by default. An admin must enable it". Editing and transitioning work items follow the connected account's permissions, unless an admin blocks Jira write.
Should we use one of the open-source Atlassian MCP servers instead?
Only after vetting it. Atlassian's security team warns an untrusted server "could be maintained by a threat actor who later applies malicious changes", and recommends pinning versions.
Sources (2), last checked 19 Oct 2026
- Atlassian Support's MCP server documentation (setup, tools, upgrade, OAuth, permissions, settings, monitoring) and security blog
- Anthropic's and OpenAI's connector documentation
Talk to an Atlassian consultant
We'll look at what you want Claude or ChatGPT to do in Jira and Confluence, and review anything already built.