MCP Atlassian Attachments

A Model Context Protocol (MCP) server for downloading Jira Cloud attachments by attachment ID to local disk, and uploading local files as attachments to a Jira issue.

Supplement to the official Atlassian MCP server. The official Atlassian MCP server covers search, issue management, and content operations but does not move attachment files between Jira and local disk. This server fills that gap.

Requirements: Python 3.13+, an Atlassian Cloud account, and an API token with at least the read:jira-work scope. Uploading also needs write:jira-work.

Authentication

Scoped tokens are recommended to limit access to exactly the permissions needed.

Note: The granular read:attachment:jira scope is not sufficient — Jira's attachment metadata endpoint (/rest/api/3/attachment/{id}) requires read:jira-work to resolve issue-level permissions. A classic (unscoped) API token also works.

Create an API token

1. Go to id.atlassian.com/manage-profile/security/api-tokens

2. Click Create API token

3. Choose "Create API token with scopes" and select read:jira-work (add write:jira-work to upload), or choose "Classic API token" for full access

4. Copy the generated token

Related MCP server: MCP Atlassian

Quick start

The fastest way to try the server is with the MCP Inspector. Set the three required environment variables and run:

From the published package (no clone needed):

ATLASSIAN_SITE_URL=https://yourorg.atlassian.net \
ATLASSIAN_EMAIL=your.email@example.com \
ATLASSIAN_API_TOKEN=your-api-token \
npx -y @modelcontextprotocol/inspector uvx mcp-atlassian-attachments

From a local clone:

ATLASSIAN_SITE_URL=https://yourorg.atlassian.net \
ATLASSIAN_EMAIL=your.email@example.com \
ATLASSIAN_API_TOKEN=your-api-token \
npx -y @modelcontextprotocol/inspector uv run mcp-atlassian-attachments

Configuration

Set environment variables or create a config file. Environment variables take priority.

Environment variables:

export ATLASSIAN_SITE_URL="https://yourorg.atlassian.net"
export ATLASSIAN_EMAIL="your.email@example.com"
export ATLASSIAN_API_TOKEN="your-api-token"

Config file (~/.config/mcp-atlassian-attachments/config.json):

{
  "site_url": "https://yourorg.atlassian.net",
  "email": "your.email@example.com",
  "api_token": "your-api-token"
}

Tools

download_jira_attachment

Returns:

{
  "product": "jira",
  "attachmentId": "439535",
  "filename": "screenshot.png",
  "mimeType": "image/png",
  "size": 496724,
  "path": "/your/output/dir/screenshot.png",
  "attachmentUrl": "https://yourorg.atlassian.net/rest/api/3/attachment/content/439535"
}

upload_jira_attachment

Returns:

{
  "product": "jira",
  "issueKey": "PROJ-123",
  "attachmentId": "600100",
  "filename": "report.png",
  "mimeType": "image/png",
  "size": 24581,
  "path": "/your/local/dir/report.png",
  "attachmentUrl": "https://yourorg.atlassian.net/rest/api/3/attachment/content/600100",
  "mediaId": "6df94659-caa7-4561-bbe3-6384ec534bba"
}

Each call adds a new attachment; Jira does not replace a same-named file. Uploads larger than the site attachment size limit fail with a 413.

get_jira_attachment_reference

Read-only. Returns mediaId for an attachment that is already on an issue:

{
  "product": "jira",
  "attachmentId": "439535",
  "filename": "screenshot.png",
  "mimeType": "image/png",
  "size": 496724,
  "attachmentUrl": "https://yourorg.atlassian.net/rest/api/3/attachment/content/439535",
  "mediaId": "6df94659-caa7-4561-bbe3-6384ec534bba"
}

After an upload you do not need this tool — upload_jira_attachment already returns mediaId, so embedding costs no extra call.

delete_jira_attachment

Needs write:jira-work. Metadata is read before the delete, so the result reports what was removed and a missing ID fails without attempting anything:

{
  "product": "jira",
  "attachmentId": "481718",
  "filename": "delete-me.txt",
  "mimeType": "text/plain",
  "size": 31,
  "deleted": true
}

There is no undo. Jira deletes the file outright, and a second call to the same ID fails with 404.

Deleting does not clean up references. If the attachment was embedded in a description or comment, the ADF media node stays exactly where it was and becomes a dangling reference — verified against a live issue. Remove the node yourself if you delete a file that was referenced.

Referencing an attachment in a description or comment

Uploading a file attaches it, but nothing appears inline in the description or a comment. Rendering it requires an ADF media node — and that node identifies the file by its Media Services UUID, not by the numeric attachment ID:

{
  "type": "mediaSingle",
  "attrs": { "layout": "center" },
  "content": [
    {
      "type": "media",
      "attrs": {
        "type": "file",
        "id": "6df94659-caa7-4561-bbe3-6384ec534bba",
        "collection": "",
        "alt": "report.png"
      }
    }
  ]
}

No Jira attachment metadata endpoint exposes that UUID, which is why this server resolves it: GET /rest/api/3/attachment/content/{id} answers with a 303 to https://api.media.atlassian.com/file/{mediaId}/binary, so reading the Location header without following the redirect yields the UUID at no download cost. The UUID exists as soon as the file is attached — it does not depend on the attachment being referenced anywhere.

Both tools therefore return mediaId. Writing the node into the issue is deliberately left to a Jira content tool such as the official Atlassian MCP server, whose addCommentToJiraIssue and editJiraIssue accept contentFormat: "adf" — this server moves files and does not edit issue content:

1. upload_jira_attachment → take mediaId from the result

2. Post a comment with an ADF body wrapping a media node built from it:

{
  "version": 1,
  "type": "doc",
  "content": [
    { "type": "paragraph", "content": [{ "type": "text", "text": "Latest run:" }] },
    { "type": "mediaSingle", "attrs": { "layout": "center" }, "content": [ "...the media node..." ] }
  ]
}

For a description, read the existing ADF, append the node, and write the whole document back — PUT /rest/api/3/issue/{key} replaces the field rather than appending to it.

Notes:

• A bare media node is not rendered; it must be wrapped in mediaSingle (single file), mediaGroup (file-card list), or mediaInline (inline chip).

• collection is "" for Jira issue attachments.

• Optional width/height on the media node set the intrinsic pixel size, and are only honoured inside mediaSingle.

• mediaId is best-effort: if the redirect cannot be read it comes back null, and the upload itself still succeeds.

Simpler alternative: v2 wiki markup

The older v2 endpoints still accept wiki markup and convert it to ADF server-side, resolving the attachment by filename — so no media UUID is needed:

POST /rest/api/2/issue/PROJ-123/comment
{ "body": "See attached: !report.png|thumbnail!" }

Jira stores that as a proper media node with the correct UUID filled in. Confirmed working on Jira Cloud as of August 2026.

Two behaviours to know about, both verified against a live issue:

• Dimensions are a fixed placeholder, not the real ones. The conversion always writes width: 200, height: 183 regardless of the image — a 64x48 and an 8x8 PNG both came back as 200x183. !file|width=300! sets the width but leaves height at 183, so the aspect ratio is wrong unless you write the ADF yourself.

• Duplicate filenames resolve to the oldest. With two attachments both named ambiguous.png, !ambiguous.png! resolved to the first one uploaded. Re-uploading under the same name to "update" an image will keep showing the old file.

Prefer mediaId when you want to be explicit: it is unambiguous with duplicate names, it lets you set true dimensions, it works with v3/ADF (which is what the official Atlassian MCP server writes), and it does not depend on a legacy path Atlassian may eventually retire.

MCP host configuration

Add the following to your MCP host's config file. The JSON is the same for Cursor (.cursor/mcp.json), Claude Desktop (claude_desktop_config.json), and Claude Code (.claude.json).

{
  "mcpServers": {
    "atlassian": {
      "command": "uvx",
      "args": ["mcp-atlassian-attachments"],
      "env": {
        "ATLASSIAN_SITE_URL": "https://yourorg.atlassian.net",
        "ATLASSIAN_EMAIL": "your.email@example.com",
        "ATLASSIAN_API_TOKEN": "your-api-token"
      }
    }
  }
}

Security

• ATLASSIAN_API_TOKEN is never logged or included in error messages.

• The computed Authorization header is never exposed in tool output or errors.

• File writes are confined to the provided output_dir.

• Filenames are sanitized to prevent path traversal, both on download and on the name sent to Jira.

• Uploads read exactly the file at file_path and send nothing else.

Tests

uv run pytest tests/ -v

Roadmap

• download_confluence_attachment_tool — Confluence Cloud uses a different API (/wiki/api/v2/) and a different identifier model. Planned for a future release.

License

MIT. See LICENSE.