On this page
Put Coworker to work on your stack.
Connect Salesforce, Slack, Jira and run your first agent in minutes.
Enterprise AI
Jira MCP Server: What It Actually Does (and How to Set It Up)
Coworker AI explains the Jira MCP server: what Atlassian's Rovo MCP Server can do, how to set it up in Claude, ChatGPT and Cursor, and its limits.
The official Jira MCP server is the Atlassian Rovo MCP Server, which lets Claude, ChatGPT, Cursor, VS Code and other AI tools search, create and update Jira work items with your existing Jira permissions. Atlassian hosts it, it has been generally available since February 4, 2026, and it works only with Atlassian Cloud, covering Confluence, Bitbucket and Loom as well.
It is not an AI model. Atlassian's FAQ calls it an integration layer and says it does not store or cache your Jira or Confluence content. It launched in beta as the Atlassian Remote MCP Server on May 1, 2025. I read Atlassian's docs and blog on October 6, 2026; sources are linked throughout.
What does the Jira MCP server do?
It turns Jira into MCP tools, grouped into permission groups that admins switch on or off. Atlassian's supported tools reference, last updated September 30, 2026, lists 52 Jira tools:
| Permission group | Tools | What they cover |
|---|---|---|
| read_jira | 27 | Work items, comments, history, worklogs, boards, sprints, versions, filters |
| search_jira | 1 | JQL search (searchJiraIssuesUsingJql) |
| write_jira | 15 | Create, edit and transition work items, comment, log time, manage sprints |
| delete_jira | 3 | Delete work items, comments and attachments (off by default) |
| manage_jira | 6 | Create and update spaces, read workflows, edit screens (off by default) |
Six Jira tools, including getJiraIssue, createJiraIssue and searchJiraIssuesUsingJql, are visible at connection time. Clients find the rest through a discover tool, a change Atlassian's changelog says reduced context window consumption by more than 50%. Rovo search (in beta) and Teamwork Graph tools add search across apps and the pull requests, builds and deployments linked to a work item.
In July 2026, Atlassian said the server handles more than 5 million tool calls a day and that nearly a third are writes, as coding agents such as Claude Code and Cursor create and update work items.
How do you set up Jira MCP in Claude, ChatGPT and Cursor?
Add Atlassian's listing from your client's directory, or point the client at https://mcp.atlassian.com/v2/mcp, then sign in with your Atlassian account and approve the scopes. Steps per Atlassian's getting started guide and each client's help docs:
| Client | How to add it | Watch for |
|---|---|---|
| Claude (web, Desktop) | Customize, then Connectors, select +, open Atlassian MCP, select Connect | On Team and Enterprise plans, an Owner adds it first under Organization settings |
| Claude Code | Run the command below, then /mcp in a session to sign in | Version 1 setups must sign in again on version 2 |
| ChatGPT | Settings, then Plugins, select Atlassian Rovo, select Connect | Company Knowledge is read-only; use the Atlassian Rovo app for writes |
| Cursor | Atlassian plugin in the Cursor Marketplace, select Add to Cursor | It installs Atlassian skills too |
| VS Code (GitHub Copilot) | Extensions view, search @mcp Atlassian, select Install | None noted |
| Other MCP clients | Server URL https://mcp.atlassian.com/v2/mcp | Gateways can add ?tools=all for a flat tool list |
claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcpSources: Anthropic, OpenAI, Atlassian's ChatGPT note and the Cursor Marketplace. Version 2 became generally available on September 8, 2026, and Atlassian's upgrade guide says version 1 connections switch to version 2 tools on March 1, 2027. For each client's plan rules, see Claude MCP and ChatGPT MCP.
How do authentication, permissions and limits work?
By default each person signs in with OAuth 2.1, and the server acts within their Jira permissions. Admins set the rest in Atlassian Administration under Rovo, then Rovo MCP server:
- Domains. Atlassian's supported AI domains, including claude.ai and chatgpt.com, are allowed by default. Admins can add domains or block the whole list, for OAuth connections only.
- Permissions. Read, Write and Search toggles per app, which take precedence over Connected Apps settings (guide).
- API tokens. Off until an admin enables them. They suit headless agents but skip the domain allowlist and log as the token's account (guide).
- Audit log. Every tool call is logged with the tool, action and user, on all tiers (guide).
- Data security policy. With Atlassian Guard Standard, admins can block MCP reads for the organization, a site or specific content. It covers only Atlassian's server and OAuth connections.
The limits: it works only with Atlassian Cloud, Atlassian's FAQ says it does not currently support FedRAMP or HIPAA requirements, calls are rate-limited per site, and some calls spend Rovo credits.
| Jira Cloud plan | MCP rate limit per site | Jira Rovo credits per user per month |
|---|---|---|
| Free | 500 calls an hour | None listed |
| Standard | 1,000 calls an hour | 25 |
| Premium | 1,000 calls an hour, plus 20 per user, up to 10,000 | 70 |
| Enterprise | 1,000 calls an hour, plus 20 per user, up to 10,000 | 150 |
Atlassian's Rovo credits page lists installing the server, single-product lookups and writes such as updating a work item as free. Enriched calls, such as Rovo search and Teamwork Graph context, usually cost 1 to 10 credits each, from an organization-wide pool shared with Rovo Chat and Agents. Extra usage billing starts December 3, 2026, and admins can cap or disable it.
Coworker
Put Coworker to work on your actual stack
Connect Salesforce, Slack, Jira and run your first agent in minutes.
Does the Jira MCP server work with Jira Data Center?
No. The official server needs an Atlassian Cloud site. A public request for a Data Center version, JRASERVER-78874, has been open since June 2025 and was still marked Gathering Interest, with 108 votes, when I checked. Jira Data Center itself reaches end of life on March 28, 2029, when it becomes read-only.
For Data Center, the main community option is sooperset/mcp-atlassian, the most-starred community Jira MCP server I found: 5,971 GitHub stars, MIT license, v0.23.1 released August 19, 2026. It supports Jira Cloud and Jira Server or Data Center 8.14 and later, exposes 98 Jira and Confluence tools, and has a read-only mode. Its README says it is not an official Atlassian product: you host and update it, and Atlassian's MCP data security policy does not cover it. None of the eight other Jira MCP repos I checked had more than 102 stars.
Where does a single-vendor Jira MCP server stop?
At the edge of Atlassian's footprint. The server is broader than Jira alone: it reaches Confluence, Bitbucket, Loom and the apps your admins connect to Rovo, and Atlassian's connector list includes Slack, Google Drive, GitHub, Zendesk and HubSpot. Salesforce and Gong were not on it when I checked.
Past that edge:
- A question like "which open Jira bugs came from accounts renewing this quarter" needs your CRM too. If the CRM is not connected to Rovo, that means a second server, and the AI client has to join the two answers itself.
- Connected servers add tool definitions to the model's context. Anthropic's help docs suggest switching Claude's connectors to On demand loading once 10 or more are active.
- Teamwork Graph covers Atlassian work and the apps connected to it. Data behind another vendor's server sits outside it, so the AI client rebuilds those links in each conversation.
- Each person signs in to each server, and each vendor's controls sit in its own admin console.
None of this is a flaw in Atlassian's server. Every vendor's own server has the same shape, including HubSpot's; the enterprise MCP guide covers rollouts across many servers.
Atlassian Rovo MCP Server vs Coworker MCP
Both are hosted MCP servers that respect source permissions. The difference is scope.
| Atlassian Rovo MCP Server | Coworker MCP | |
|---|---|---|
| What it is | Atlassian-hosted remote server for Atlassian Cloud | Server that exposes your company's data and workflows to any MCP-compatible AI tool |
| Jira | 52 Jira tools, Jira Cloud only | One of 50+ connected tools |
| Beyond Jira | Atlassian apps, plus apps connected to Rovo | 50+ tools, including Salesforce, Slack, GitHub, Gong and Zendesk |
| Context behind it | Teamwork Graph | OM2 organizational memory (for large organizations) |
| Sign-in | OAuth 2.1, or API tokens if an admin enables them | Existing Coworker login and SSO |
| Admin controls | Domains, Read/Write/Search toggles, IP allowlists, audit log, data security policy | MCP access enabled per customer by your admin, with usage visibility |
| Cost | Included with Atlassian Cloud; enriched calls use Rovo credits | Not published; book a demo |
| Vendor benchmark | Agents grounded in Teamwork Graph context gave 44% more accurate answers with 48% fewer tokens | 9x lower token cost and 64% faster in the best case; preferred 84.5% of the time in blind review* |
| Best fit | Work that lives in Jira and Confluence Cloud | Questions that span Jira and tools outside Atlassian |
*Statistically significant benchmarks comparing Coworker MCP vs. Claude Native Tooling: Claude with Coworker MCP and OM2 against Claude with its native connectors (method), 100 tasks across 7 categories, with Jira among the connected sources. The 9x and 64% figures are the best case (retrieval-heavy categories, with Coworker Learning); across all 7 categories the same setup was 75.5% cheaper and 45.6% faster. Atlassian's post does not describe its method, so the numbers are not directly comparable.
What should you check before connecting Jira MCP?
- Confirm you are on Jira Cloud and on version 2.
- Keep OAuth for people. Use API tokens only for headless agents, on a least-privilege service account.
- Start with Read and Search, add Write later, and leave delete_jira and manage_jira off.
- Decide whether to keep Atlassian's default AI domains or allow only the clients you approve.
- Check IP allowlists, since some AI tools call from their own IP ranges, and allow *.atlassian.net so interactive Jira widgets load.
- Require approval for writes: Claude Team and Enterprise owners can set write tools to Needs approval.
- Watch the audit log (filter for Rovo MCP User Actions) and Rovo credit use before extra usage billing starts.
One rollout tip from Atlassian's README: put your cloudId, Jira project key and a maxResults of 10 in an AGENTS.md file to cut discovery calls and tokens. For the threat side, read Atlassian's MCP risk note and the MCP security guide.
Where Coworker MCP fits
If you only need Jira inside Claude, ChatGPT or Cursor, use Atlassian's server. Atlassian builds and hosts it, and connecting costs nothing beyond Rovo credits for enriched calls. If you already run Rovo with connectors for most of your stack, try Rovo search first.
Coworker MCP is for questions where Jira is one of many systems, including ones outside Rovo's list such as Salesforce and Gong. One connection gives Claude, Claude Code, ChatGPT, Cursor, Windsurf or any MCP client access to 50+ tools, including Jira, Slack, GitHub and Zendesk, and all MCP queries respect your existing permission model. An admin enables MCP access, people sign in with their existing Coworker login and SSO, and setup typically takes under 30 minutes.
For large organizations, it is backed by OM2. "You can think of OM2 as a context graph," Coworker's CEO wrote in the launch post. Per the OM2 page, documents, messages and tickets are decomposed into atomic facts linked to entities and relationships. The launch post adds that source permissions are enforced on traversal, and that once Coworker solves a class of retrieval task, such as constructing the right Jira JQL, it learns the path for next time. More in what a context layer is and the Glean MCP comparison.
To decide, test both on the same questions from your own Jira and nearby tools. Coworker will run its benchmark on your data and hand you a Context Impact Report, or you can book a demo.
Frequently asked questions
Does Jira have an official MCP server?
Yes. Atlassian builds and hosts the Atlassian Rovo MCP Server, which connects Claude, ChatGPT, Cursor, VS Code and other MCP clients to Jira, Confluence and other Atlassian Cloud apps. It launched in beta on May 1, 2025 and became generally available on February 4, 2026.
Is the Jira MCP server free?
Connecting is included for Atlassian Cloud customers, and Atlassian lists installation, single-product lookups and writes as free. Enriched calls, such as Rovo search, usually cost 1 to 10 Rovo credits each, drawn from allowances in paid Jira plans. Extra usage billing starts December 3, 2026.
How do I install Jira MCP?
Add Atlassian's listing from your AI client's connector directory, or add the server URL https://mcp.atlassian.com/v2/mcp, then sign in with your Atlassian account. In Claude Code, run claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp, then /mcp to sign in.
Does the Atlassian MCP server work with Jira Data Center?
No. It requires an Atlassian Cloud site, and the public request for a Data Center version was still gathering interest in October 2026. Community servers such as sooperset/mcp-atlassian support Jira Server and Data Center 8.14 and later, but you host and secure them yourself.
Can Jira MCP create and update tickets?
Yes. The write_jira group can create, edit and transition work items, add comments, log time and manage sprints. Delete and admin tools are off by default, and admins can turn off writes in Atlassian Administration.
Is Jira being phased out?
No. Jira Cloud continues, and Atlassian says it is shifting its full focus to cloud. Self-hosted Jira Data Center is what ends: sales to new customers ended on March 30, 2026, and it becomes read-only on March 28, 2029.
How is Coworker MCP different from the Jira MCP server?
Atlassian's server goes deep on Atlassian Cloud apps, plus apps connected to Rovo. Coworker MCP gives AI clients one connection to 50+ tools, including Jira, Salesforce and Slack, with OM2 organizational memory for large organizations. If Jira is all you need, Atlassian's server is the simpler choice.
Related reading
Ready to get started?
Put Coworker to work inside your actual stack
Connect Salesforce, Slack, Jira, whatever you use, and run your first agent in minutes.