jira2mcp

A published stdio MCP server for Jira Cloud. Requires Python 3.13+.

Use jira2mcp with any MCP client that can launch a stdio server. It does not support Jira Server or Data Center, and does not provide dedicated issue assignment, issue delete/archive, sprint/board/epic, or admin-configuration operations.

Install and configure

Install uv, then configure your MCP client to start the server:

{
  "mcpServers": {
    "jira": {
      "command": "uvx",
      "args": ["jira2mcp"]
    }
  }
}

For Claude Code:

claude mcp add jira -- uvx jira2mcp

The server uses stdio. Its only launch option is:

--credentials-file PATH  Explicit path to a Jira Cloud JSON credentials file

For example:

{
  "mcpServers": {
    "jira": {
      "command": "uvx",
      "args": ["jira2mcp", "--credentials-file", "~/.config/jira-cloud.json"]
    }
  }
}

Authentication

Set environment credentials:

export JIRA_URL="https://yourcompany.atlassian.net"
export JIRA_USER="you@company.com"
export JIRA_API_TOKEN="your-api-token"

Or pass an explicit file, which takes precedence over environment credentials:

{
  "url": "https://yourcompany.atlassian.net",
  "username": "you@company.com",
  "api_token": "your-api-token"
}

There is no default credentials-file path and no JIRA_CREDENTIALS_FILE environment variable. Create a token in your Atlassian account; never commit, print, or expose it in prompts or shared configuration.

Tools

All tools use the jira_* namespace.

Identity, reads, and transitions

Metadata and saved filters

Issues, comments, and links

Attachments and worklogs

The server also provides the data://jira/link-types resource and the jira_jql_syntax prompt.

Usage and safety

Descriptions and comments accept Markdown and are converted to Atlassian Document Format (ADF). jira_read returns selected Jira fields unchanged, including ADF, as structured content and a compact JSON text fallback; it does not format or truncate the response. jira_run_filter returns the same search-shaped result as jira_search after resolving the filter's JQL. jira_download_attachment provides structured/raw-friendly output, while jira_attachment remains available for its original simple download surface.

jira_read requires both issue_key and a non-empty native fields array, for example jira_read(issue_key="PROJ-123", fields=["summary", "description"]). Each array item is one field key, ID, or endpoint-supported selector; do not use comma-separated items or surrounding whitespace. It has no raw mode because it always returns structured Jira data. Selectors such as *all, *navigable, or negative selectors can still return broad responses, so request only what is needed.

Before a create or edit, call jira_fields for the target project and issue type. Before a transition, link, comment update/delete, attachment deletion, or worklog mutation, read the current state and use exact IDs or names. Attachment uploads must stay within the server working directory. Downloads must stay within advertised MCP roots, or the server working directory when roots are unavailable. All reads and writes remain subject to the configured Jira account's permissions.

Field catalog

jira_list_fields returns exactly one Jira /field/search page. Its start_at default is 0, max_results default is 20, and it accepts optional project_key, query, native field_ids arrays, and field_types arrays containing "system" and/or "custom". Normal text is concise (field name and canonical ID). With raw=True, it returns the complete Jira page envelope, including values, startAt, maxResults, total, isLast, and any other Jira properties as structured content and a JSON text fallback.

To continue, retain every filter and use startAt + len(values) as the next start_at; stop when Jira reports isLast. Jira can cap the requested page size, so use its returned metadata. The endpoint is documented for classic projects. project_key only supplies Jira project context/access filtering: it does not identify fields applicable to an issue type, Create screen, or Edit screen. Use jira_fields for issue-type, create-screen, and edit-screen metadata.

Changelog history

jira_changelogs retrieves every Jira GET page for an issue before returning one helper-owned structured envelope, {"issue_key": "<KEY>", "changelogs": [...]}. Its optional timestamp bounds are client-side filters applied after complete retrieval with created_at_or_after <= created < created_before; they do not limit Jira's GET requests. Its optional native field_ids array retains only changelog items whose raw fieldId exactly matches a canonical field ID, dropping history events with no retained items.

jira_changelogs normally returns every retained event. Supplying result_max_results enables local event pagination; result_start_at defaults to 0 and requires result_max_results when nonzero. Jira's complete history is still fetched before timestamp and field-ID filters and the result slice. Paged raw output adds helper-owned result_page metadata; it is not Jira server pagination.

Use jira_changelogs_by_ids only for IDs already known from that history. It uses Jira's distinct known-ID POST endpoint and accepts the same native field_ids array filter. Jira controls the returned order. This known-ID operation has no result pagination, so request fewer IDs when a smaller response is needed.

Normal text is condensed and server-truncated at 30,000 characters. With raw=True, both changelog tools return the untruncated helper-owned envelope as MCP structured content plus JSON text fallback, not untouched API pages. Complete histories and raw results may be large and can still be clipped by an MCP client or harness.

Search pagination and raw output

Each jira_search or jira_run_filter invocation returns exactly one issue page; jira_search fetches one issue page, while jira_run_filter resolves its saved filter and then fetches one issue page. max_results is per page, defaults to 20, and is capped at 50. To continue, pass the non-empty response nextPageToken as the next request's next_page_token; the names intentionally differ. Keep the same JQL (or saved filter ID), fields, and page size for every call. For example:

jira_search(jql="project = PROJ", max_results=20, fields=["summary"], raw=True)
jira_search(jql="project = PROJ", max_results=20, fields=["summary"], next_page_token="<nextPageToken>", raw=True)

Use raw=True from the first page. Raw mode returns the complete API-shaped page as both MCP structured content and a JSON text fallback, including nextPageToken and arbitrary requested fields. Normal formatted text has a fixed issue view and is server-truncated at 30,000 characters; raw mode is not server-truncated, but an MCP client or harness can still clip its result. If a current page was externally clipped, continuing with its token can skip issues that were not displayed.

Field selection is whole-field projection, not nested projection or redaction. Omitting fields requests all seven defaults: summary, status, assignee, priority, issuetype, created, and updated. In particular, assignee can include nested identity, email, and avatar data allowed by the Jira account. Request explicit fields when that data is not wanted. Arbitrary fields are available in raw mode; normal formatted text remains the fixed issue view.

Continue whenever nextPageToken is non-empty, including when an intermediate page has no issues. Stop only when the token is absent or empty, not when total is reached or reported. Tokens expire after seven days; restart from the first page when one expires. jira_run_filter resolves the saved filter on every call, so do not edit the filter while paging. Jira result-set changes, including a changed saved filter, can otherwise produce duplicates or omissions.

For the published Jira CLI, see the jira2cli guide. Contribution and maintainer guidance is in the repository contributing guide.

License

MIT