Jira mcp

MCP server for Jira Server/Data Center (REST API v2). Lets an AI assistant (Claude Code, Claude Desktop, ...) read and write your Jira directly.

Requirements

• A Jira Server/Data Center account (username + password).

• Node.js 18+.

Related MCP server: Jira MCP Server

Quick start

Use Claude Code CLI. Run the command in a regular terminal, not inside a Claude Code session. Pick the block for your shell, since each one continues lines differently.

bash / zsh (macOS, Linux, Git Bash or WSL on Windows):

claude mcp add g-jira-mcp \
  -e JIRA_HOST="https://jira.company.com" \
  -e JIRA_USERNAME="your_username" \
  -e JIRA_PASSWORD="your_password" \
  -- npx -y g-jira-mcp@latest

PowerShell (Windows):

claude mcp add g-jira-mcp `
  -e JIRA_HOST="https://jira.company.com" `
  -e JIRA_USERNAME="your_username" `
  -e JIRA_PASSWORD="your_password" `
  -- npx -y g-jira-mcp@latest

Command Prompt (cmd.exe): put the whole command on one line, or end each line with ^ instead of \.

Or manually add to .claude/settings.json (or claude_desktop_config.json):

{
  "mcpServers": {
    "g-jira-mcp": {
      "command": "npx",
      "args": ["-y", "g-jira-mcp@latest"],
      "env": {
        "JIRA_HOST": "https://jira.company.com",
        "JIRA_USERNAME": "your_username",
        "JIRA_PASSWORD": "your_password"
      }
    }
  }
}

Restart Claude Code/Desktop after editing the config.

To check that it connected, run /mcp inside Claude Code, or claude mcp list in a terminal.

Official setup guides:

• Claude Code: Connect Claude Code to tools via MCP explains claude mcp add, the local, project and user scopes, and the shared .mcp.json file.

• Claude Desktop: Connect to local MCP servers shows, with screenshots, how to open claude_desktop_config.json from Settings, Developer, Edit Config, and where to find the logs when a server does not start.

Environment variables

To discover custom field IDs on your instance:

curl -u user:pass https://jira.company.com/rest/api/2/field | jq '.[] | select(.name | test("story|point|start"; "i")) | {id, name}'

Tools

Notes

• Jira Server uses plain text for descriptions — no ADF format.

• get_ticket omits unset fields rather than printing placeholder text. Key, Summary, Status and Assignee are always present (Assignee shows "Unassigned" when empty); other fields appear only when set. Pass include_comments=true to append the ticket's comment thread; comments are off by default to keep output small and save tokens. The same Jira request fetches everything, so including comments costs no extra API call. Comments are capped at 20 most recent; if a ticket has more, the header reads Comments (20 most recent of 45): so the caller knows older comments exist and can open the ticket in Jira to see them. Each comment line starts with its ID (#10001), which update_comment and delete_comment take.

• get_ticket lists attachments as name (#id) in an Attachments: line whenever the ticket has any. Pass include_images=true to also receive image attachments (PNG, JPEG, GIF, WebP) as MCP image content, so the caller can see screenshots referenced as !image.png! in the description or comments. Images are off by default because each one costs far more tokens than the text. At most 5 images are returned by default (raise it with image_limit, up to 20), each up to about 3.7 MB and about 7.5 MB in total, so the response stays inside Claude's vision limits (5 MB base64 per image on Bedrock and Google Cloud, 32 MB per request). Images are sent as-is; the model API downscales large ones itself. Each image costs one extra download request; an Images: section maps [image N] to its file name, reports each skipped image with its reason (too large or total size cap reached), reports failed downloads, and when image_limit is hit, says how many images were left out. Raising image_limit brings back only images that hit the count limit; images skipped for size will not reappear.

• get_comment returns a single comment as #id [date] author: body. Pass include_images=true to fetch and return only the image attachments that comment embeds via !filename.png! or !filename.png|width=…! markup (by contrast, get_ticket returns all images on the ticket). Fetching attachments costs one extra request only when the comment embeds something. Images follow the same per-image and total size limits as get_ticket, up to 20 max. Use get_comment instead of get_ticket for ?focusedCommentId= links, or when only a single comment matters.

• search_tickets output is adaptive: empty columns (no value anywhere) are dropped entirely, and constant columns (same value on every row, when there are 3+ rows) are stated once in the header as All: Status=In Progress and removed from the table. When columns are dropped, the header also notes which ones were empty across all results (e.g. Unset for every row: Priority, Parent, Start Date) so the caller can tell "no ticket has a due date" from "this tool doesn't return due dates". KEY and Summary are always kept. This keeps results focused and token-efficient. The header's Found N issue(s) (showing M) reports when the result was truncated — raise max_results above the default 25 to see more.

• create_ticket and update_ticket both accept an assignee parameter (Jira username as a string, sent as {name: assignee}). In create_ticket, an empty assignee value is ignored. In update_ticket, pass assignee="" to unassign.

• create_ticket applies original_estimate in a follow-up PUT rather than in the create payload: Jira Data Center answers a bare 500 Internal server error when timetracking is present in POST /issue, while the same value applies cleanly as an update afterwards. The ticket key is reported even if that second call fails, with a warning to set the estimate via update_ticket — so a failed estimate never sends you back to create a duplicate.

• duedate is a standard field (YYYY-MM-DD); "Start date" is a custom field, configurable via JIRA_START_DATE_FIELD.

• update_ticket's epic_key sets the Epic Link field (a custom field, configurable via JIRA_EPIC_LINK_FIELD, default customfield_10001) so the ticket shows under the epic's "Issues in Epic" panel — not the same as link_issues, which only creates a generic Linked Issue (e.g. "Relates"). Epic Link can only be set on standard issue types (Story/Task/Bug), not on Sub-tasks or Epics themselves.

• search_tickets uses JQL syntax, e.g. project = PROJ AND status = 'In Progress'.

• add_attachment uploads one local file and returns its attachment ID. Pass an absolute file_path: a relative path resolves against the server process working directory, which the MCP client chooses. To show an uploaded image inline, reference it as !filename.png! in a comment or description. The server can read any file its user can read, so check the path before you approve the call.

• move_to_sprint takes a sprint name or numeric ID. A name is looked up among the active and future sprints of the scrum boards of the first ticket's project (Agile API, /rest/agile/1.0); if it matches none, or several, nothing is moved and the error lists the open sprints with their IDs. To list a sprint's tickets, use search_tickets with JQL such as sprint = "PROJ Sprint 12" or sprint in openSprints().

• transition_ticket resolves the transition ID automatically, matching either the transition's own name ("Resolve Issue") or the status it lands on ("Done"). Common aliases map onto whatever the workflow actually offers (Closed/Resolved/Complete → Done, Reopen → Re-Open, Todo → To Do, Cancelled → Won't Do), so the same call works across workflows with different status names. Exact matches win, then aliases, then a substring fallback. When nothing matches, the error lists every valid option for that issue as Transition -> Target Status; the tool description cannot list them because they vary per issue and workflow.

• When Jira rejects an unknown issue type (create_ticket, update_ticket, search_tickets) or link type (link_issues), the error also lists the valid names: the project's issue types when the project is known, every issue type for a JQL search, and every link type.

• update_ticket only changes the fields you pass; omit a field to keep its current value. Pass assignee="" to unassign, and due_date="" or start_date="" to clear a date. implementation_notes appends to the description. Jira's REST API ignores a parent change on an existing sub-task, and rejects or ignores an issue type change to or from a sub-task type, so update_ticket makes those through the same Move/Convert wizards the Jira UI uses: sub-task to standard type, standard type to sub-task (needs parent_key), between sub-task types, and a new parent for a sub-task. It stops, changing nothing, when the new type's workflow lacks the ticket's current status (the error lists the statuses to pick from in the Jira UI), needs a required field, or would delete field values. A change between two standard types that REST does not apply still needs the Jira UI "Move" action.

• log_work advertises one canonical name per work type (code, deploy, design, management, meeting, misc, operation, qa, req, research, translation) and per activity (correct, create, review) so its schema stays small; every alias still works as input (coding, dev, testing, ops, requirement, other, and the rest). "fix" is an activity alias (Correction), not a work type. The form shows Type of Work and Type of Activity as two independent dropdowns, so any work type pairs with any activity. activity is required when work_type is set, and comment is required whenever work_type or activity is set, because the WorklogPRO form rejects an empty Work Description. When the form rejects a submit, the error carries the form's own message and nothing is logged. If neither is set, logs via plain REST (no WorklogPRO form). If work_type or activity is set and the WorklogPRO form is unavailable, nothing is logged and the call fails (safe to retry); it never falls back to REST, which cannot set those attributes. A failed form submit is an error and is never retried, because the worklog may already exist. Start times are interpreted in the Jira server timezone (configurable via JIRA_TIMEZONE).

• get_ticket with include_worklogs=true appends the ticket's worklogs, each line starting with its ID (#20001), which update_worklog and delete_worklog take. This costs one extra request (GET /issue/{key}/worklog), capped at the 20 most recent like comments. log_work returns no ID when it goes through the WorklogPRO form, so this is where to find it. update_worklog edits time, start and comment over REST only; to change Type of Work or Type of Activity, delete the worklog and log it again with log_work.

• generate_release_notes groups tickets by type into Features / Improvements / Bug Fixes / Other.

• Register the server as g-jira-mcp and use that key in every project. The key becomes the tool prefix (mcp__g-jira-mcp__get_ticket), so a project that registers it under a different name exposes different tool names — an agent carrying the habit of one name into a project configured with the other gets No such tool available.

• Transient Jira failures (429, 500, 502, 503, 504) are retried up to twice with a 250ms/500ms backoff. Only GET, PUT and DELETE are replayed — a POST that returned 500 may already have created the comment, worklog or transition, so it fails fast instead of risking a duplicate.

• All logs go to stderr; stdout is reserved for the MCP protocol.

Example prompts

• "Search tickets in project PROJ that are In Progress"

• "Create a Story in PROJ titled 'Release notes v2.0' due 2026-08-01"

• "Update PROJ-234, set the assignee to namcp and add label BugFix"

• "Add a comment to PROJ-234: 'Review done'"

• "Generate release notes for fix version v2.4 in project PROJ"

Troubleshooting

• 401/403: recheck JIRA_USERNAME/JIRA_PASSWORD and whether the account can access the project.

• Connection/timeout: verify JIRA_HOST format (starts with https://, no trailing /), and whether VPN/internal network is required.

• Start date not saving: confirm JIRA_START_DATE_FIELD matches your instance (see the discovery command above).

• Error logs: every failed tool call is written to stderr with a timestamp and the tool name, and the startup line shows the server version. On Windows, Claude Code keeps this in %LOCALAPPDATA%\claude-cli-nodejs\Cache\<project>\mcp-logs-<server name>\ (for example mcp-logs-g-jira-mcp).

• Reporting a bug: open an issue at https://github.com/ameforge/atlassian-mcp/issues with the server version, the tool you called and the matching log lines. Logs can contain ticket content and internal URLs, so remove anything sensitive before you post. Nothing is sent anywhere automatically.

Development

pnpm install
cp .env.example .env   # edit with your credentials
pnpm build             # bundle to dist/index.js via rsbuild
pnpm test

Release from the repo root with pnpm release:jira; see docs/RELEASE.md.

Support

Questions or issues? Email PNam.

If this project helps you, consider buying me a coffee:

Created by PNam