How to Work with the Jira MCP Server Day-to-Day: Setup, Capabilities, Prompts, and Fixes

[#1]

An AI assistant that can read your Jira tasks behaves differently from one that cannot. It stops asking you to paste context and starts working from your real sprint, your real backlog, and your real specs. The connection that makes this possible takes only a few minutes to set up.

However, extracting tangible value can take longer. You need to know what can be achieved with a Jira MCP server, how to phrase a request to make it safe, and what to do when the answer looks wrong. This article follows that order. We start with what the MCP server is and what it can read, then move on to setup and example prompts. We also share three practical workflows worth building, along with fixes for the problems that project management teams encounter most often.

Definition

A Jira MCP server is a connector that lets an AI assistant read and write your Jira data through the Model Context Protocol, an open standard for linking AI tools to external systems. 

Does Jira have an MCP server?

Yes, there is an official Jira MCP server provided by Atlassian. It's called the Atlassian Rovo MCP Server, and it covers more than Jira: it also connects your AI assistants to Confluence, Compass, Jira Service Management, and Bitbucket Cloud. Atlassian does not offer a separate MCP server for Jira alone. When people search for the Jira MCP server, what they really need is the Rovo MCP Server. For simplicity, in this article, we will still call it Jira MCP Server, although it's worth bearing in mind that it connects to other Atlassian tools as well.

Jira MCP Server provides you with a working connection between your Atlassian data and your AI assistant, such as ChatGPT, Gemini, and Claude. Ask a question about your sprint, and the assistant will answer from your actual work items. Ask it to create a bug, move a story to In Review, or add a comment, and the change lands in Jira.

Here are the tools and platforms that can be connected to the Jira MCP server:

• Claude and Claude Code
• ChatGPT
• Mistral
• VS Code
• Cursor
• GitHub Copilot
• Codex
• Docker MCP Toolkit
• AWS
• Azure
• Google Cloud
• Figma
• Postman
• Writer
• Lovable
• Devin
• Resolve

This list keeps growing as Atlassian regularly adds new connectors.

The server is free to use with your Atlassian Cloud site - Jira Cloud, JSM Cloud, and so on. However, you might need to pay separately for the AI assistant you use - for example, if you are on a paid ChatGPT plan. Jira MCP server is hosted and fully maintained by Atlassian, so you don't need to do anything yourself in this regard.

Apart from the official Jira MCP server, there are also community-built MCP servers - open-source projects that you need to host yourself. Unlike the official Jira MCP server, most of them work with Jira Data Center too, not just Jira Cloud. I compared both options in my article What Is Jira MCP? Key Concepts, Capabilities, and Use Cases, so you might want to start there if you need a more detailed picture. In this article, we will focus on Atlassian's official MCP server.

The official Rovo MCP server is widely adopted across various industries. Atlassian reports that over 1 million users work with Atlassian Rovo MCP every month, making over 5 million tool calls every business day. Nearly a third of those calls are writes, which means assistants are creating and changing work, not only reading it passively. Atlassian also found that about 44% of all Rovo MCP users are not working in software development teams.

What the Jira MCP server can do for you

"Every Friday, I ask my AI assistant the same question: What did the team actually finish this week? It reads the work items we closed in our product space, and the Smart Checklist progress on everything still open. The second half is the part I need. On the board, a task with 9 out of 10 steps done looks the same as a work item with 2 out of 10 steps completed. But they result in very different Mondays. So it's great that the Jira MCP server allows me to pull progress information in such granular detail".

Oleksandra Sokol

Product Lead at TitanApps 

Teams put the Jira MCP server to work in various ways. Some of them replace a few clicks, and others replace an afternoon of digging through spaces, pages, and repositories. Here are the key capabilities of the Jira MCP server you should know about:

• Find and read work. Your AI assistant can search across spaces, sprints, and work item types without you writing JQL queries. Ask something like "What did the design team ship this month?" and you get a list back in seconds. For a single item, the assistant can pull its description, comments, linked pages, and the change history in one reply. From there, you can keep asking follow-ups in the same thread.
• Create and update work. Everything you can change on a work item, your assistant can change on your behalf. That covers creating work items with fields already filled in from a plain description, editing priorities and assignees, adding comments, logging time, moving items through workflow statuses, and linking work items. Bulk work is where this saves the most time. A page of meeting notes could become eight work items in one request - all with one prompt.
• Understand your Jira setup. Your assistant can also read how your Jira is configured, including the spaces you have access to, the statuses in a workflow, the link types available, and the account IDs behind usernames. These lookups will probably rarely interest you directly, but they are what keep the other answers correct. An assistant that has read your workflow will not try to move a story into a status you do not have.
• Work across Atlassian tools. The same conversation can reach Confluence, Bitbucket Cloud, Compass, and Jira Service Management. You can create an epic and its linked spec page together, or read a Confluence brief and use what it says to fill in the work items that implement it. Teamwork Graph is the underlying data layer that connects these pieces. It maps the relationships between a work item, the page describing it, and the pull request closing it. All this allows you to automate actions and workflows across multiple tools.
• Connect to tools outside Atlassian. An AI tool can hold multiple MCP connections at once, so a single request can include data from Jira and a completely different system. For instance, this allows you to pull a work item, find its GitHub pull request, and update both in one prompt. This is also the ground floor for AI agents, which often need several connected systems before they can plan multi-step work.

Here's what Anna Kohnen, VP of Partnerships at Figma, says about using Jira MCP server: “MCP connectors in Figma have become a core part of product development workflows for entire teams. Whether it’s a product manager pulling in a Confluence doc to spin up a prototype in Figma Make, or a designer referencing Jira tickets to explore ideas with the Figma design agent, teams can build with shared content on the multiplayer canvas while keeping updates in sync with their Atlassian tools.“

In total, the Jira MCP server exposes more than 60 individual actions across these products (in documentation, these actions are called "tools"). You don't need to learn their names to use them, since your AI assistant picks the right ones from your request automatically.

What data the Jira MCP server can access

The capabilities in the previous section rest on a specific set of data, and that data comes from three places. Most of it is your Jira and the Atlassian apps connected to the same server. The rest arrives from tools outside the Atlassian ecosystem and from third-party Marketplace apps that keep their content on your work items.

Work item and Atlassian app data

Your assistant can reach the fields on a work item, its comments and worklogs, its links, and the transitions currently available on it. Add to that the results of any JQL or natural language query you run against a space you can open. Content from the connected Atlassian apps comes through the same connection. Confluence pages arrive with their body content and comments, Bitbucket Cloud contributes repositories and pull requests, and Jira Service Management adds ops alerts and on-call schedules.

External tool data through Teamwork Graph

Most teams keep at least part of their context outside Atlassian, and that part is reachable too. Teamwork Graph is Atlassian's data layer that maps relationships between work items, pages, repositories, and people. It supports more than 100 out-of-the-box connectors. Once your organization connects Slack, GitHub, Salesforce, or Workday, that data becomes available in the same conversation as your Jira work items. Adoption varies from one organization to another, so check which connectors are actually in place before you build a prompt around one.

Note that the connectors feeding the Teamwork Graph have to be configured by your Atlassian administrator first. Using the Teamwork Graph tools with the Rovo MCP server is free while they are in open beta, though Atlassian has said this will be billed in Rovo credits later.

Marketplace app data

External apps can store their content in custom fields and issue properties, which are accessible to AI assistants through the Jira MCP server. The server can read them the way it reads a status or an assignee, which means that app data reaches your AI assistant with no extra setup on your side. 

Smart Checklist for Jira is an example of this. It is a third-party process management solution that lets you break down complex processes into actionable ToDos without subtask overhead. In practice, this means that Smart Checklist enables you to add feature-rich checklists to your Jira work items. This is convenient for organizing work on multi-step tasks or for managing quality criteria, such as the Definition of Done. 

Checklist content lives in Jira issue properties and in the Checklists custom field, and the completion information is stored in the Smart Checklist Progress field. As a result, your AI assistant can access information on checklist progress alongside everything else on a work item. This allows you to better understand your team's progress and organize your work.

Marketplace apps can also contribute to Teamwork Graph directly, although that option is newer and few apps use it so far.

How to set up the Jira MCP server

Every AI tool has its own flow for this, which makes the process look more varied than it is. However, underneath, the same four things happen in the same order. That holds for a chat assistant, an IDE, or any of your other development tools. Once you have done it once, the next tool will only take a few minutes.

Check what you need first

Access to an Atlassian Cloud site is the only hard requirement, and no admin approval is needed to get started. One detail catches teams out, though. The first person from your organization to connect determines the app scope for everyone who follows. For example, if that person has no Confluence access, colleagues may later find Confluence missing from their own connection. When you are not sure whether you are first, ask your admin before you start.

Point your AI tool at the server

Every connection uses the same Atlassian endpoint:

https://mcp.atlassian.com/v1/mcp/authv2

How you supply this piece of information depends on the tool. Some list Atlassian in a connector gallery or an extension marketplace, so you install it and never see the URL. Others read it from a config file, in JSON for most clients and in TOML for Codex. A few older clients cannot reach a remote server directly and need the mcp-remote proxy, which requires Node.js v18 or later.

Sign in and approve access

A browser tab opens, you log in with your Atlassian account, and you approve the access the AI tool is asking for. Authentication runs on OAuth 2.1 by default, and the token it issues is bound to that client. From then on, the assistant acts with your permissions and nothing more.

API token authentication is the alternative, mainly for headless setups. It is also currently required for Jira Service Management and Bitbucket Cloud actions, and an organization admin has to switch it on first.

Confirm the connection works

One prompt is enough to test it. Ask something narrow with a verifiable answer, such as "List the Jira spaces I have access to." A healthy response names your real spaces. A failed one either reports no Atlassian access or answers in generalities without naming anything. The second case usually means the request never reached Jira.

Step-by-step instructions for individual AI tools are provided in the FAQ on Jira MCP Server Setup at the end of the article.

How to structure your Jira MCP prompts to get a real result

The gap between a vague answer and a useful one usually comes down to four pieces of information. Weak prompts are missing at least two of them. Here are the elements a proper prompt should include.

Scope: name the space, sprint, or time window

Without those, the assistant has to guess how wide to search, and wide searches are slow and expensive.

Weak: "Show me our bugs."

Strong: "Show me bugs in the WEB space created in the last 14 days, highest priority first."

Fields: say which data the assistant should read

This matters most for fields that come from Marketplace apps, since the assistant will not check them unless you ask.

Weak: "Which work items are nearly done?"

Strong: "List work items in the current sprint with their status and their Smart Checklist Progress value."

Action: state whether to read, create, transition, or link

Be explicit about what exactly you want the AI assistant to do. Ambiguous verbs are what produce unexpected writes.

Weak: "Handle PROJ-512."

Strong: "Read PROJ-512 and summarize the open questions in the comments. Do not change anything else."

Guardrails: tell the assistant what to check before it writes

Tell the assistant what to check before acting, and cap how much it returns. Search before creating is the single most useful rule here.

Weak: "Create a bug for the login error."

Strong: "Search the WEB space for an existing bug about login errors on mobile. If you find one, add a comment with these details. If not, draft a new bug and show it to me before saving."

Behind the scenes, your AI tool converts requests like these into a JQL query and a series of API calls. That is why precision in the prompt turns into precision in the search. Your wording does not have to match Atlassian's current vocabulary, though. For example, plenty of people still use old terminology - they can ask for "new issues" or to look up a "Jira issue by key" instead of saying "Jira work item" or "Jira ticket". The assistant maps all of that to work items without complaint.

A starter prompt library for using Jira MCP Server (grouped by job)

These prompts are also good examples of how the Jira MCP server can be used day-to-day by agile teams.

Find and retrieve information from Jira

• "Show me all open work items assigned to me in the PAY space, grouped by status."
• "Find the work item about the OAuth refresh bug in the PAY space, even if the summary doesn't use those exact words."
• "Pull PROJ-922 with its description, acceptance criteria, and all comments."
• "Which work items in the WEB space have been in progress for more than 10 days, and who owns them?"
• "List everything in the current sprint that has changed status since Monday."

Report and spot patterns in your Jira work

• "List work items in the current sprint with their status and Smart Checklist Progress, and flag any at In Review with open checklist steps."
• "Show me stories closed in the last two sprints where the Definition of Done checklist is not completed."
• "Show how many work items each team member completed last week and how many they still have open in the current sprint. Sort the team members by the number of remaining work items."
• "Compare this sprint's bug count to the previous three, and group by component."

Create Jira work items and Confluence pages in bulk

• "From these meeting notes, draft eight work items in the MOB space and a Confluence page summarizing the decisions and open questions. Show me everything before creating anything."
• "Create an epic called Q3 Mobile Redesign in the MOB space, then add the five child stories listed below."
• "Read this Confluence spec and create one work item per section, linked to the parent epic. Check for duplicates first."

Move work through the workflow

• "Transfer PROJ-512 to In Review and add a comment linking the pull request."
• "Transition every work item in the v4.2 release that has an approved pull request to Ready for QA. Show me the list and ask for confirmation before transitioning any of them."
• "Assign PROJ-200 to Marta, set priority to High, and leave a comment explaining why, based on this Slack thread."

Pull data across Confluence, Bitbucket, and other sources

• "Find the launch brief in Confluence for the Spring Campaign and pull out the key dates."
• "Which work items in the current sprint have linked GitHub pull requests that are still open?"
• "Read the retro page from the last sprint and check which of its action items became work items."
• "Summarize what shipped in the MOB space this week, including the linked Bitbucket pull requests."

You can also describe your role and typical tasks to your AI assistant and ask it to brainstorm ways you could use the Jira MCP server in your daily work.

Three Jira MCP workflows worth setting up first

Almost everything above can also be done inside Jira. The workflows worth building are the ones where an assistant does something the Jira interface cannot. It reads several sources at once, compares them, and forms a judgment that you then check. All three below are read-heavy, which makes them safe to try early.

Workflow 1: gather work item context before you start coding

Scenario: An engineer picks up a story in the morning and needs the full picture before writing any code. The context is spread across the work item, a Confluence spec, a checklist, and last month's related pull request.

Requirements: The Jira MCP server connected, Confluence access available through the same connection, and Bitbucket linked if you want the git history included.

What happens: The assistant reads the work item and its comments, opens the linked spec, checks the checklist steps, and looks at related pull requests. It then drafts an implementation plan and lists what is still ambiguous. The engineer reviews the plan, corrects it, and only then starts working.

Why it beats Jira alone: Jira can show you all four sources, one at a time. It cannot tell you that the spec says one thing and the work item description says another. Catching that contradiction before the software development work starts is where the time is saved. Apart from that, you save time by not having to switch between several tools.

Workflow 2: turn a Confluence spec into a Figma prototype

Scenario: A product manager has a feature spec in Confluence and a handful of work items describing what it should do. The team is ready to see the idea rather than read about it, so the next step is a clickable prototype.

Requirements: The Jira MCP server connected to Figma, which is a supported client, and access to the Confluence space holding the spec.

What happens: The assistant pulls the spec and the linked work items into Figma, then Figma Make turns them into a working prototype. The requirements arrive as text that the design agent can act on, so the flows and states in the prototype match what the work items actually ask for. Designers can then continue to reference those same work items as they explore alternatives on the canvas.

Why it beats working in Jira alone. Jira holds the requirements and cannot do anything visual with them. Rovo does not help here either, since the work happens on the Figma canvas, outside Atlassian. The connection is what lets the spec cross that boundary without anyone copying it across by hand.

Workflow 3: clear a stale backlog across several spaces

Scenario: A backlog has grown for two years across three spaces, and nobody wants to spend a day reading it. Grooming keeps getting postponed because the first hour is pure sorting.

Requirements: Jira MCP server connected

What happens: The assistant finds work items that are stale, unassigned, unestimated, or likely duplicates, then returns a triage list grouped by problem. You approve the closures and merges, and the assistant applies them in a second pass.

Why it beats Jira alone: A JQL filter can find items untouched for six months, but it cannot notice that four of them describe the same login bug in different words. Reading for meaning across hundreds of summaries is the part that does not fit into a filter.

When using the Jira MCP server is not the right approach

An MCP server broadens what your assistant can do in Jira, but it does not fit every workflow. Some tasks are easier to handle in the Jira interface, and others belong in dedicated automation or reporting tools. Here are the cases where a different approach will serve you better:

• Simple, one-off actions on a work item already in front of you. If the change takes one click, describing it in chat will take longer than just doing it. Save the MCP for multi-step tasks, bulk changes, and questions that would otherwise need JQL.
• Scheduled, recurring automation. Jira Automation is purpose-built for this. If you want the same status change every Friday, set it up there once and let it run. The same goes for a subtask that should appear whenever a parent is created. An assistant is better for ad-hoc requests than for background jobs.
• Simple questions related to Atlassian tools. Rovo Chat is already inside Jira and Confluence, so opening an external assistant for the same question adds an extra step. If you want a summary of a work item you have open, or an answer about what a Confluence page says, ask there instead. Rovo Chat can also draft work items from a description with the /create-work-items command. The MCP server earns its place when the answer depends on something outside Atlassian, when the question is complex and requires integrating data from multiple sources, or when you want to have control over which AI models you use.
• Visual planning on boards and timelines. Dragging cards across a Kanban board, reshaping a timeline, or laying out a dashboard are visual decisions. You can describe them in text, but the feedback loop would be slower than just pointing and clicking, so such tasks are faster to do inside Jira.

 

Troubleshooting: 5 common Jira MCP server problems

Most of what goes wrong falls into a handful of typical patterns. Here is how to recognize them and what to do in each case.

Why does creating a work item fail on required fields?

You ask for a new work item and get an error instead of a confirmation, usually naming a field you may not recognize. 

Why it happens. Sometimes your Jira space requires that a new work item contain fields that the AI assistant did not fill in. Standard fields usually go through without any trouble, while required custom fields can become the sticking point, especially fields with selection lists that need an exact option value.

How to fix it. Ask your AI assistant to create a work item with standard fields only, and once it's done, ask it to update the custom fields in a second request. Atlassian keeps improving this, so this problem might go away in the future.

Why am I hitting rate limits so quickly?

Why it happens. One prompt can fan out into many API calls behind the scenes. A request phrased as "list everything" burns through the allowance long before you notice.

How to fix it. Narrow the scope with a space key and a time window. Ask for summaries instead of full content, and split bulk writes into smaller batches. If your team runs into the ceiling regularly, check where you are in Jira's points-based quota, since the allowance scales with your plan. For more about this, please refer to the official documentation about Jira Cloud's points-based quota.

Why does Jira work but Bitbucket does not?

Why it happens. Bitbucket Cloud and Jira Service Management actions require API token authentication, and using OAuth alone does not allow you to access them.

How to fix it. An organization admin needs to enable token authentication in the Rovo MCP server settings. Start with your admin and confirm if the correct settings are enabled. If token authentication is already on, check that the Bitbucket workspace is linked to your Atlassian organization, because older workspaces were often set up independently.

Why does the assistant reference the wrong space?

Why it happens. When you do not specify which space you want it to access, your AI assistant has to guess. Typically, it tends to pick whatever space it saw last in the conversation.

How to fix it. Add your default space key and your account ID to your AI tool's system instructions or rules file. After that, when you ask your AI assistant something like "show me my open work items", it pulls information from the correct Jira space (Jira project) without you spelling it out each time.

Why is the AI assistant creating duplicate work items?

Why it happens. Your AI assistant does not check for similar existing work items unless you ask it to. A prompt that says "create a task" will create it, even when a matching work item is already present in the backlog.

How to fix it. Add a search-first instruction to any prompt that creates something, and ask for confirmation before bulk writes. Reviewing the proposed items before they are saved takes seconds and removes the need for a cleanup entirely. If you often create new work items with an AI assistant, you can add this to your instruction files in the AI client project or create dedicated skills for this purpose.

Four habits that turn Jira MCP Server setup into a team capability

• Keep a shared prompt library. A request that returns exactly the right answer should land on a team page the same day, grouped by use cases so people can find the closest match.
• Standardize your issue-tracking structure. Prompts written by one person only work for everyone else when field names and work item types follow the same pattern across spaces.
• Agree on which AI tools people use. Development tools like VS Code, Cursor, and Claude Desktop each run a different LLM, so results vary even with identical prompts.
• Bring your admin in early. API token authentication for Bitbucket and Jira Service Management actions requires an organization admin, and admins also control which external tools can connect.

For more information about this topic, please see my earlier article on the Atlassian Community, Jira MCP Explained: What It Does, What It Doesn't, and How to Use It. 

FAQs about Jira MCP server setup

How do I connect the Jira MCP server to ChatGPT?

Open Settings, go to Plugins -> Browse Plugins and look for Atlassian Rovo in the Plugins Directory. Add it, and a browser window opens for the Atlassian sign-in. The connector is available on Plus, Pro, Business, Enterprise, and Edu plans; on Business and Enterprise workspaces, an admin may need to enable it first.

How do I add the Jira MCP server to VS Code with GitHub Copilot?

Open the Extensions view, type @MCP Atlassian in the search field, select the Atlassian MCP server from the gallery, and install it. VS Code handles the endpoint for you and opens the browser for authentication.

To add it by hand, open the command palette and choose the command for adding an MCP server. Select HTTP as the connection type and paste the server URL:

https://mcp.atlassian.com/v1/mcp/authv2

Give the server a name, then open Copilot Chat in agent mode and ask it something about your Jira to confirm the connection.

How do I connect Claude Desktop to the Jira MCP server?

Go to Settings, open Extensions, select Browse extensions, then Plugins. Search for Atlassian and install the plugin. Claude Desktop will open your browser for the OAuth sign-in, and the Atlassian actions will become available in your next conversation.

Alternatively, you can add the Jira MCP Server directly to your config file:

{

  "mcpServers": {

    "atlassian": {

      "url": "https://mcp.atlassian.com/v1/mcp/authv2"

    }

  }

How do I set up the Jira MCP server in Cursor?

Go to the Cursor Marketplace and find the Atlassian plugin for Cursor with MCP. From that page, you will be able to add the Jira MCP server with one click and sign in with your Jira credentials.

You may need to restart Cursor afterward. 

How do I set up the Jira MCP server in Claude Code?

Run one command in your terminal:

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

This registers the server in your local scope, so it loads in the project where you ran it. Then, open a new Claude Code session and run /mcp to authenticate and confirm that Atlassian appears in the list.

How do I set up the Jira MCP server in Codex?

On Codex Desktop, you can install Jira MCP through the UI: Open Plugins or Connectors, find Atlassian Rovo, and install it from there.

If you are not on desktop, you can install by running this command:

codex mcp add atlassian --url https://mcp.atlassian.com/v1/mcp/authv2

Then, start a new session so Codex picks it up, and complete the authentication steps.

Jira MCP server FAQ: general questions

What is the full list of Jira MCP server tools?

The server exposes more than 60 actions (tools) across Jira, Jira Service Management, Confluence, Compass, and Bitbucket. They cover reading and searching work items, creating and updating them, running transitions, adding comments, working with Confluence pages, and reading repositories. Atlassian keeps the complete table in its Rovo MCP Server documentation, and the list is updated regularly.

Does the Jira MCP server work with Jira Data Center?

Not the official one. The Atlassian Rovo MCP Server is Cloud-only and works with sites on the atlassian.net domain.

Teams on Data Center can use open-source community MCP servers instead, such as mcp-atlassian. These are usually distributed as Docker images and run on your own infrastructure. The functionality is broadly similar, but the trade-off is that hosting, patching, and security become your responsibility.

Can I use the Jira MCP server without giving the assistant write access?

It's not possible to do that by flipping a single switch, since the OAuth scopes are granted when you connect. However, many AI tools let you disable individual actions for a connected server, so you can turn off the ones that create or update. In the example below, all write/delete permissions for Atlassian Rovo in Claude are set to "Needs Approval".

Most clients also ask for confirmation before a write runs, which gives you a checkpoint every time. 

The strongest control available is your Atlassian account's access scope itself. Your AI assistant inherits your Jira permissions, so if you have an account with read-only project permissions, your AI assistant will only have read access too.

Does the Jira MCP server see data from Marketplace apps?

Yes, when the app stores its data in custom fields or issue properties (work item properties). Both are standard Jira storage locations, so the Jira MCP server can easily read them.

Smart Checklist for Jira works exactly this way, which is why prompts about checklist progress allow your AI assistant to fetch precise information. The same pattern applies to other apps that expose their data through those fields

[Sagar Mahajan]

Thanks for writing this details article, @Olga Cheban _TitanApps_