Jira MCP Server

About

The Jira MCP server is Atlassian's official Remote MCP Server, giving AI assistants like Claude and Cursor direct, enterprise-grade access to Jira Software project management through natural-language interactions. Powered by Atlassian's Teamwork Graph and hosted on Cloudflare infrastructure, it requires no local process to run — authentication is handled via OAuth 2.1, making it the most secure way to connect AI to Jira in corporate environments. With this MCP server, product managers, engineers, and team leads can ask their AI to create and update Jira issues, transition ticket statuses through workflow stages, search with JQL (Jira Query Language), summarize sprint progress, view open epics and their child issues, retrieve assignee workloads, and bulk-triage backlogs. AI assistants can connect sprints to related Confluence documentation through Atlassian's graph layer, giving richer context for planning and retros. Enterprise customers including AT&T, NVIDIA, and Pfizer use Atlassian's MCP integration in production. Connect from Claude Desktop via Settings > Connectors, or add it to Claude Code with: `claude mcp add --transport http atlassian https://mcp.atlassian.com/v1/mcp/authv2`. Cursor and Windsurf users add the remote URL to their MCP config file. No install command needed — it's a fully hosted remote MCP server.

Installation

claude mcp add --transport http atlassian https://mcp.atlassian.com/v1/mcp/authv2

Connecting Jira over the Atlassian Rovo MCP Server

1. 1.Add the remote server to Claude Code

   One command, no package download. Claude Code opens a browser for the OAuth 2.1 consent screen the first time you call a tool; after that the session is remembered.

   shell

   claude mcp add --transport http atlassian https://mcp.atlassian.com/v1/mcp/authv2

2. 2.Or paste the config directly, for any HTTP-capable client

   This is the exact .mcp.json Atlassian ships in the repository root, so it is the config to copy rather than one reconstructed from a screenshot. Cursor, VS Code and ChatGPT also have one-click install links in the README that encode the same URL.

   json

   {
     "mcpServers": {
       "atlassian": {
         "type": "http",
         "url": "https://mcp.atlassian.com/v1/mcp/authv2"
       }
     }
   }

3. 3.Clients that only speak stdio need the mcp-remote proxy

   If your client cannot open a streamable-HTTP connection itself, it runs mcp-remote locally to bridge to the hosted endpoint. Atlassian lists Node.js v18+ as a prerequisite for exactly this path — that requirement is about the proxy, not about the server.

   shell

   npx -y mcp-remote https://mcp.atlassian.com/v1/mcp/authv2

4. 4.Let the first consent be an admin, or the rest of the team is blocked

   The Rovo MCP app is not a Marketplace install. It is registered just-in-time the first time somebody on your site completes the OAuth 2.1 (3LO) consent flow, and that first user has to hold access to every product the MCP scopes request. Get that wrong and later users hit "Your site admin must authorize this app". After the first install, a user with only Jira (or only Confluence) can consent for their own product. The app then shows up under Connected apps in Atlassian Administration, where it can be reviewed or revoked.

5. 5.Pin the cloudId and project key so the model stops probing

   Every session otherwise opens with getAccessibleAtlassianResources and a project lookup before it does the thing you asked for. Atlassian publishes this AGENTS.md block to short-circuit those calls; the maxResults line matters as much as the IDs, because an unbounded JQL search will happily return enough issues to blow the context window.

   markdown

   ## Atlassian Rovo MCP
   
   When connected to atlassian-rovo-mcp:
   - **MUST** use Jira project key = YOURPROJ
   - **MUST** use cloudId = "https://yoursite.atlassian.net" (do NOT call getAccessibleAtlassianResources)
   - **MUST** use `maxResults: 10` for ALL Jira JQL search operations.

The Jira tools, by permission group

Access is granted per permission group, not per tool — an org admin turns on read_jira, write_jira or search_jira, and each tool inherits its group. That is why "it can read my issues but not comment" is a normal, deliberate state rather than a bug. atlassianUserInfo and getAccessibleAtlassianResources are common to every product.

What people use it for

Which one should you use?

One hosted server covers several products, so "which Jira MCP server" is usually really a question about deployment and auth.

Categories

Works With

Frequently Asked Questions

Related Guides