github sooperset/mcp-atlassian v0.24.5

55 minutes ago

🔒 Security

This release fixes security issues found in an internal audit of the server. Upgrade if you run the HTTP transports (sse, streamable-http), use JIRA_PROJECTS_FILTER, CONFLUENCE_SPACES_FILTER or JIRA_INTERNAL_ONLY_PROJECTS, or let callers choose the Atlassian URL with request headers. Details are in each advisory. The fixes are in #1886.

Project and space filters

  • GHSA-f5qp-f5cw-2r8f: tool arguments that become part of a request path (comment, queue, request type, issue type, version, field and context IDs, and the page move position) can no longer address a resource outside the allowlist.
  • GHSA-vhwv-5pg8-hh56: with a filter set, empty scoped arguments are refused, and changelog items, Confluence comment bodies, templates and space lookups stay inside the allowlist.

Server-side requests and operator settings

  • GHSA-cvhg-qx9x-fv7g: Jira attachments are downloaded only from the Jira instance the request is for, and JIRA_SSL_VERIFY=false / CONFLUENCE_SSL_VERIFY=false apply only to the configured origin. Operator policy settings (JIRA_INTERNAL_ONLY_PROJECTS, timeouts, markup translation) now also apply to requests that authenticate with URL and token headers.

HTTP transports

  • GHSA-qx98-vrcc-c5w9: when requests can be served with the operator's own credentials (ALLOW_GLOBAL_CRED_FALLBACK or IGNORE_HEADER_AUTH), the server checks the Origin header on every address and the Host header on loopback.
  • GHSA-xqr5-x7vp-hmg9: Jira and Confluence calls no longer block the server for every other caller. They run in worker threads with a limit per upstream site, circuit breaker state is kept per site, and large comment, worklog and paged reads are bounded.
  • GHSA-3jvr-xc3p-xhgr: converting Jira and Confluence markup takes time in proportion to the size of the text.

Hardening

  • mcp 1.30.0. Picks up the fix for GHSA-fmmv-w9g8-j3gc in the MCP SDK (#1799).
  • Attachment downloads stop at the size limit (#1881), and one tool call returns at most 50 MB of attachment data inline.
  • Token characters stay out of debug logs (#1881).
  • Confined file opens do not follow symbolic links, and one dynamic client registration is limited in size.

⚠️ Behavior changes to check before upgrading

Most deployments need no change. Two settings may need attention:

  1. ALLOW_GLOBAL_CRED_FALLBACK or IGNORE_HEADER_AUTH with a browser client or a same-host proxy. A request with an Origin header gets 403 unless the origin is local or listed in FASTMCP_HTTP_ALLOWED_ORIGINS. A request that arrives on a loopback address with a non-local Host gets 421 unless the host is listed in FASTMCP_HTTP_ALLOWED_HOSTS. Clients that send no Origin header are not affected.
  2. JIRA_SSL_VERIFY=false or CONFLUENCE_SSL_VERIFY=false with URL headers. Certificate checks are off only for the configured origin. A request that names another host in its URL header gets this only when that host is under MCP_ALLOWED_URL_DOMAINS. Redirect targets and other hosts verify certificates.

Other changes:

  • IDs are validated. jira_edit_comment, jira_get_queue_issues, jira_get_request_type_fields, jira_create_customer_request, jira_get_create_fields and jira_update_version refuse an ID that is not numeric. jira_get_field_options refuses a field_id with characters other than letters, digits and underscores.
  • confluence_move_page accepts append, above, below, before and after, with the same meaning on every auth path. On Cloud OAuth above and below used to fail.
  • Large reads stop early and say so. confluence_get_comments and confluence_get_inline_comments read at most 200 comments per collection. When a read stops early, list responses end with one {"truncated": true, "message": ...} item and object responses carry "truncated": true and a hint. Jira worklog reads stop at 5000 entries.
  • Mentions. One converted Confluence page or comment looks up at most 50 different users. Mentions past that show the stored ID.
  • With JIRA_PROJECTS_FILTER or CONFLUENCE_SPACES_FILTER set: an empty string for a project, issue, space or content argument is refused. Changelog output drops link, parent, epic, project and sprint changes that name issues outside the allowlist. Confluence comment bodies are converted from storage format, so include and listing macros show as placeholders. A template with no space must be a global template.
  • With JIRA_INTERNAL_ONLY_PROJECTS set: comment writes on other projects make one more issue lookup and are refused if it fails.
  • ATLASSIAN_CIRCUIT_BREAKER_THRESHOLD now opens the breaker for one upstream site at a time. JIRA_FETCHER_MAX_WORKERS (default 8) is a limit per upstream site for both Jira and Confluence.
  • Diagram blocks. Fenced mermaid and plantuml blocks in Confluence Markdown are stored as code blocks, and draw.io processing is off. The earlier mermaid macro pointed at an attachment that was never uploaded.
  • Jira text conversion. Bracketed text right after a wiki link no longer breaks the link: See [docs|http://x] and [PROJ-1] now becomes See [docs](http://x) and PROJ-1 now.
  • Dynamic client registration refuses a client_name over 256 characters or metadata over 16 KiB.

🐛 Bug Fixes

Server / Transport

  • The configured timeout applies to every request. Requests that the server sends directly through the HTTP session, such as attachment uploads, had no timeout. One stalled upload could block a shared streamable-http server until the process was killed. These requests now use JIRA_TIMEOUT and CONFLUENCE_TIMEOUT (#1795). Thanks to @JosXa for the diagnosis and the fix.

Jira

  • jira_create_issue accepts its own parameters inside additional_fields. Passing components, assignee or another named parameter in additional_fields raised got multiple values for keyword argument before any request was sent. The tool now merges them, combines components from both places, accepts components by id, and rejects only values that contradict each other (#1796, #1834). Thanks to @vkpeter for the report and @star7js for the fix.
  • expand=renderedFields returns the rendered content. jira_get_issue and jira_search now return a rendered_fields object keyed by field id (#1733, #1739). Thanks to @kenneth-burgener-sndk for the report and @vishalhabib99 for the fix.

Confluence

  • Cloud OAuth: labels, comments, the space page tree, page history and page diff work again. Atlassian removed the Confluence v1 content endpoints from the OAuth gateway (410 Gone). confluence_get_comments returned an empty list, and the others failed. These reads now use the v2 API (#1598, #1373, #1783). Thanks to @jasandov for the fix and @Code-by-AB for the report.
  • A comma in an attachment comment no longer fails the upload on Cloud. Confluence Cloud reads the comment field of a new upload as a comma-separated list, so such a comment returned HTTP 400. The comment is now set right after the upload and is stored as written. Upload errors include Confluence's reason (#1803, #1857, #1883). Thanks to @generalpurpose123 for the report and @TejasLamba2006 for the diagnosis.
  • confluence_upload_attachments reports failed files. The summary counts uploaded and failed files, and a batch in which no file was uploaded returns success: false (#1883).
  • Page history and space page reads return the stored body unchanged with convert_to_markdown=false. They used to run the body through the preprocessor, which flattened user mentions and could change macros. Writing that output back lost the original structure (#1725). Thanks to @zifter.
  • Macro parameter values no longer leak into Markdown. A panel rendered as #deebff#f4f5f71Title and a status as BlueIN PROGRESS, because parameters of macros without a dedicated handler were kept as text (#1750, #1838). Thanks to @HeeyunCho for the report and @TejasLamba2006 for the fix.
  • A page that was created is no longer reported as failed. Confluence Cloud can answer 404 for a new page for a moment after it was created. confluence_create_page then reported a failure, and a second attempt made a second page. The read after a create is now tried again, and if it still fails the error says the page was created and gives its ID (#1885).
  • A refused page move is reported as an error. With an API token, Confluence can refuse a move with a status below 400. The tool then reported success for a page that had not moved (#1833). Thanks to @davekim917.
  • Pages with many code macros convert faster. The converter re-serialized the whole page once per code macro (#1880).

Docs and CI

  • Installation instructions for GitHub Copilot in VS Code (#1836). Thanks to @frazar.
  • Every GitHub Action is pinned to a commit, workflows are audited with zizmor, and an unused workflow is removed (#1882, #1884).

Contributors

Thanks to @JosXa, @star7js, @vishalhabib99, @jasandov, @TejasLamba2006, @zifter, @davekim917 and @frazar, and to @vkpeter, @kenneth-burgener-sndk, @Code-by-AB, @generalpurpose123 and @HeeyunCho for the reports.

Full Changelog: v0.24.4...v0.24.5

Don't miss a new mcp-atlassian release

NewReleases is sending notifications on new releases.