Atlassian Rovo MCP Server

Atlassian's official MCP server connects an AI agent to Jira, Confluence, Jira Service Management, Bitbucket Cloud, Compass and Loom on Atlassian Cloud. It is hosted by Atlassian at https://mcp.atlassian.com/v2/mcp, signs in with OAuth 2.1 or an API token, and honors each user's existing permissions. In Claude Code, add it with one claude mcp add line. We checked the README on 31 October 2026.

Quick facts

What it lets an agent do

The README sorts the uses into four verbs: summarize and search, retrieve and review (Loom recordings), create and update, and automate repetitive work such as turning meeting notes into work items.

Sample prompts from the README show the range:

• Jira: "Find all open bugs in Project Alpha." "Make five Jira issues from these notes."
• Confluence: "Summarize the Q2 planning page." "Create a page titled 'Team Goals Q3'."
• Compass: "What depends on the api-gateway service?"
• Loom: "Summarize this Loom walkthrough and list the action items."
• Combined: "Move PROJ-456 to 'In Review' and add a comment that the PR is up."

The last one is the most useful for developers. An agent that has just opened a pull request can move the ticket and leave a comment without you switching tabs.

Two versions, and why it matters

New setups should use v2. The README says v2 exposes more tools and more products. The older v1/mcp/authv2 and v1/mcp endpoints remain supported, and existing v1 connections start to use v2 tools automatically.

That automatic switch has a side effect. Clients that cannot finish authentication after the move may need to clear cached client IDs or .well-known credentials. If a connection that worked last month now fails, start there.

If you configured a custom client against https://mcp.atlassian.com/v1/sse, change it. The README warns that after 30 June 2026 that endpoint is no longer supported, and says to use /mcp, preferably the v2 URL.

How the tools are exposed

Most MCP servers send every tool definition at connection time. Atlassian does something different. The README says the server exposes a small set of the most-used tools, marked Primary, and lets the client discover the rest on demand. That keeps context free for the task, and new tools can appear without a reconnect.

There is an override. If you use an MCP gateway or need every tool up front, add ?tools=all to the URL: https://mcp.atlassian.com/v2/mcp?tools=all. It returns a paginated flat tool list.

The README does not list tool names. It points to Atlassian's Supported tools page for the full reference and the scope each permission group needs. Admins grant access by permission group, and each tool inherits the access of its group.

Permission groups

Conditions the README states:

• delete_jira and manage_jira are disabled by default. An admin must enable them.
• Bitbucket Cloud tools work only if the workspace is linked to an organization.
• Jira Service Management tools need API token authentication, and an admin must have enabled it.
• Loom tools work only for workspaces linked to an Atlassian site.
• Teamwork Graph tools can read linked pull requests, builds and deployments from connected services. For GitHub for Atlassian, full access follows the user's GitHub permissions, and limited access follows their Jira permissions. Azure DevOps, GitLab, Jenkins and Spinnaker follow the limited model.

Install it in each agent

The README has a table of vendor commands for Claude Code, Codex and others, plus one-click buttons. Lines for Antigravity, Gemini CLI and OpenCode are our adaptation.

Claude Code

From the README:

claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp

Then run /mcp inside a session to authenticate. Claude users on the desktop app and Claude.ai can also install the Atlassian connector from the Claude directory, and the repo ships a Claude Code plugin manifest.

Codex

From the README:

codex mcp add atlassian --url https://mcp.atlassian.com/v2/mcp

Our addition: run codex mcp login atlassian if the sign-in does not start by itself. The Codex desktop app can install "Atlassian Rovo" from Plugins or Connectors.

Cursor

The README links an Add to Cursor button and an Atlassian plugin in the Cursor marketplace, whose skills come with the plugin. By hand, our adaptation of Cursor's remote format:

{
  "mcpServers": {
    "atlassian": { "url": "https://mcp.atlassian.com/v2/mcp" }
  }
}

VS Code

The README's button writes {"url": "https://mcp.atlassian.com/v2/mcp", "type": "http"} under the name Atlassian-Rovo-MCP. You can also search @mcp Atlassian in the Extensions view.

Gemini CLI, Antigravity and Windsurf

The README lists Google Gemini CLI as a supported client and the repo ships a Gemini extension manifest. Gemini CLI is now for Code Assist enterprise customers, and individuals moved to Antigravity CLI on 18 June 2026. For Antigravity, our adaptation uses its serverUrl field in ~/.gemini/config/mcp_config.json:

{
  "mcpServers": {
    "atlassian": { "serverUrl": "https://mcp.atlassian.com/v2/mcp" }
  }
}

For Windsurf the README says to add a custom server in Cascade with serverUrl set to the v2 URL..

OpenCode

OpenCode starts the OAuth flow on its own when a remote server answers 401. Our adaptation:

{
  "mcp": {
    "atlassian": {
      "type": "remote",
      "url": "https://mcp.atlassian.com/v2/mcp",
      "enabled": true
    }
  }
}

Local clients and mcp-remote

For tools that only run local stdio servers, the README names the mcp-remote proxy and says Node.js 18 or newer. See MCP clients for how each agent names these fields.

Sign-in options

OAuth 2.1 opens a browser and needs you to sign in and approve access. It is the default for people at a keyboard.

API tokens are for headless use, such as backend systems and automations, and they are required for Jira Service Management tools. Two kinds are supported:

An organization admin must first turn API token authentication on, under Atlassian Administration, Rovo, Rovo MCP server, Authentication. You also need a scoped token. If the admin has disabled token auth, clients must use OAuth.

Service account keys are the better choice for a shared automation, because they have their own identity. A personal token ties the agent's actions to one person.

Security and permissions

The server enforces its own controls, per the README: traffic over HTTPS with TLS 1.2 or later, OAuth 2.1 or API tokens, data access that respects each user's permissions in every connected product, and IP allowlist rules honored for tool calls.

Admin controls worth knowing:

• Permission groups. Grant or revoke by group. Keep delete and manage off unless you have a specific need.
• Domain controls. The Rovo MCP server settings page lists which external AI tools and domains may connect. Atlassian-supported domains are allowed by default; you can add or block. These apply to OAuth connections.
• IP controls. If you use IP allowlisting, MCP requests must come from allowed addresses. A common error is "You don't have permission to connect from this IP address".
• Revocation. Admins can review or revoke the MCP app's access. Users can revoke their own authorizations from their profile settings.
• Audit log. Each tool use is recorded. In Atlassian Administration, go to Insights, Audit log, and filter for Rovo MCP User Actions, or search "MCP".

For people:

• The agent acts as you. Whatever you can edit in Jira, it can edit. Writes appear in history with your name.
• Ticket and page text is untrusted. A Jira description or a Confluence page can contain text aimed at the model. An agent that reads customer-submitted tickets and can also write elsewhere is exposed to prompt injection. Keep such agents on read permissions.
• Bulk create is easy to overdo. "Make five issues from these notes" works. So does "make five hundred". Ask for a preview first.
• Service desk data is sensitive. JSM tickets may hold customer details. Limit the token's scopes to the projects the agent needs.

Cutting tool calls and tokens

The README offers a tip we would take. Add a short block to your AGENTS.md that fixes the Jira project key, the Confluence space ID and the cloud ID, so the agent skips discovery calls. It also suggests setting maxResults or limit to 10 for every JQL and CQL search. Both save tokens and speed up each task. It also points to default Atlassian skills in the repo's skills/ folder, and says Cursor's marketplace plugin includes them.

Limits and known issues

• Cloud only. The README describes an Atlassian Cloud bridge.
• Auth gaps by product. JSM needs an API token, and Compass needs OAuth. You cannot use one method for everything.
• Admin gates. API tokens and delete or manage tools are off until an admin turns them on. A blocked setup is often a policy setting, not a bug.
• Version churn. The v1 to v2 move can break cached credentials, and the SSE endpoint was announced as ended after 30 June 2026.
• Linked workspaces. Bitbucket and Loom tools need the workspace linked to an organization or site.
• Open issues. The repo showed 108 open issues when we checked. We did not read them..

Alternatives

GitHub MCP and this server pair well: the agent reads the ticket, makes the change, opens the pull request and updates the ticket. The MCP servers hub lists more.

Which agents it suits

• Claude Code and Codex fit the ticket-to-pull-request loop.
• Cursor and VS Code are the README's first two one-click installs, aimed at referencing issues and logging work from the codebase.
• Claude and ChatGPT suit managers who want summaries of projects and goals.
• Teams with strict IT gain from the admin controls, which are more detailed than most servers offer.

Our take

We did not connect to an Atlassian site for this page. This is from the README and the links it gives.

Atlassian has done the work an enterprise needs. Permission groups, domain controls, IP rules, audit logs and a service-account path are all there, and delete tools are off by default. Few servers match that.

The cost is complexity. Two auth methods with different product coverage, two endpoint versions and an on-demand tool model mean you spend setup time on policy. For a solo developer, the OAuth path in Claude Code takes a minute. For a company, plan an hour with an admin.

We would start with read and search groups, and add write once people trust the agent. Keep delete off. And put the project key and cloud ID in AGENTS.md, since the README itself suggests it.

A short setup checklist

1. Confirm your site is Atlassian Cloud.
2. Add the v2 URL with the line for your agent.
3. Sign in with OAuth.
4. Ask for a read-only task, such as open bugs in one project.
5. Put the project key, space ID and cloud ID in AGENTS.md.
6. Ask an admin about API tokens only if you need headless or JSM access.
7. Check the audit log to see your own tool calls.

Frequently asked questions

What is the Atlassian MCP server URL?

https://mcp.atlassian.com/v2/mcp.

Does it work with Jira and Confluence?

Yes, along with Jira Service Management, Bitbucket Cloud, Compass and Loom, on Atlassian Cloud.

How do I add it to Claude Code?

Run claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp, then /mcp to sign in.

Can an admin control what the agent can do?

Yes, by permission group, plus domain, IP and API token settings.

Is the legacy SSE endpoint still available?

Not after 30 June 2026, per the README. Use /v2/mcp.