Jira / Confluence Team Lead

An MCP server and an Angular dashboard that answer the questions a team lead actually has, and can publish the answer to Confluence:

1. What's in this sprint, and is it going to land?
2. How did each developer do over the last few sprints? (real changelog-derived cycle/lead time, not a guess from current status)
3. What's not ready for the next refinement session, and who should be in the room?
4. Who has gone quiet? — active work with no comment for over a day.

Built and tested against a personal Atlassian Cloud site first; the target corporate instance is a .env change, not a code change.

Layout

src/jira_confluence_mcp/   the report engine (metrics, rendering, Atlassian client)
  server.py                MCP server        — stdio, for Claude
  api.py                   HTTP API          — for the dashboard
frontend/                  Angular dashboard
scripts/demo_server.py     run the dashboard on synthetic data, no Jira needed

The MCP tools and the dashboard call the same report builders, so a number shown in the UI and a number quoted in chat can't drift apart.

See it working without Jira

Two terminals, no credentials required:

.venv\Scripts\python.exe scripts\demo_server.py

cd frontend && npm start

Then open http://localhost:4200. The data is invented, but it flows through the real changelog parsing and metric code.

Setup

python -m venv .venv
.venv\Scripts\python.exe -m pip install -e ".[dev]"

Copy .env.example to .env and fill in the base URL, your Atlassian email, and an API token from https://id.atlassian.com/manage-profile/security/api-tokens.

Then register the server (.mcp.json in this repo already does it for Claude Code — adjust the absolute paths if you move the project):

claude mcp add jira-confluence -- "C:\DEV\Jira Confluence MCP\.venv\Scripts\python.exe" -m jira_confluence_mcp.server

First call should always be check_connection — it verifies both products, prints the resolved defaults, and tells you which custom field it picked for story points.

The dashboard

.venv\Scripts\python.exe -m jira_confluence_mcp.api

cd frontend && npm start

The dev server proxies /api to the backend on port 8000, so there's no CORS to configure. Four screens:

• Overview — completion against time elapsed, points, at-risk count, a changelog-derived burndown, load per person, and a filterable issue table.
• Developers — the multi-sprint table, one row per developer per sprint. Sort any column, filter by developer or sprint, export to CSV. This is the thing Confluence's static tables can't do.
• Refinement — what's not ready, why, and a copyable attendee list.
• Quiet work — active issues with no recent comment, per person.

Every screen has a Publish to Confluence button that writes the same report as the MCP tool would. The browser sends the report kind, never markup — the server regenerates the storage format itself, so the API can't be used to inject arbitrary content into a Confluence page.

Reports are cached for 120 s so the dashboard doesn't hammer Jira; each page's Refresh button bypasses the cache.

Node version: the CLI is pinned to Angular 21 because Angular 22 needs Node ≥ 22.22.3 and this machine has 22.17.0. Upgrade Node first if you want to move to 22.

Tools

Tool	What it does
check_connection	Auth check + resolved config. Run this first on a new instance.
list_sprints	Sprints on a board, so you can find sprint ids. Falls back to listing boards if none is configured.
get_sprint_issues	Issues in a sprint, filterable by status name or status category. Key, summary, assignee, story points, status.
get_developer_sprint_history	2–5 sprint rollup per developer: completed, carried over, avg time in progress, cycle time, lead time.
get_refinement_readiness_report	Stories missing an estimate or DOR/DOD labels, plus a suggested attendee list.
get_sprint_health	One-call sprint overview: totals, breakdowns, burndown. Powers the Overview screen.
get_stale_issues	Active work with no comment for over N days — who's gone quiet.
publish_confluence_page	Create or overwrite a page from storage-format XHTML.
check_confluence_capabilities	Cloud vs Data Center, and whether the Table Filter app is available.

Every report tool returns rows (structured), markdown (for reading), and — with include_confluence_storage=true — confluence (storage format). Pass publish=true to write it straight to a page; re-running overwrites the same page by title.

Metric definitions

These are the decisions baked into metrics.py. Worth agreeing with the team before the numbers get used in a retro.

• Completed — the issue was in a Done-category status at sprint close (completeDate, else endDate, else now for an active sprint). Judged at sprint close, so an issue finished the week after the sprint ended counts as carried over for that sprint. Never evaluated past now, so an active sprint is measured against today rather than its future end date.
• Carried over — everything in the sprint that was not completed.
• Re-opened work — an issue that was Done and moved back out is not completed. Closed → re-opened → closed again reports the final close.
• Time in progress — calendar hours in In Progress-category statuses, summed across visits, capped at completion or sprint close.
• Cycle time — first entry into an In Progress-category status → completion.
• Lead time — issue created → completion.
• Grouping — by the issue's current assignee. Reassigned work is credited to whoever holds it now; the changelog has the data to split it by holder if that turns out to matter.

Status categories (new / indeterminate / done) are read from the instance rather than hard-coded status names, so a workflow that calls it "Development" instead of "In Progress" still works. Any status the instance does not report is surfaced in the report's warnings.

Confluence tables

Storage-format tables are static — not sortable or filterable. Options, in the order worth trying on a target instance:

1. table_style="table-filter" — wraps the table in the Table Filter, Charts & Spreadsheet macro. Run check_confluence_capabilities first; reading the app inventory needs admin rights, so a null result means "ask an admin", not "not installed".
2. Confluence Cloud's database content type — behaves like an embedded spreadsheet, but cannot be created through the REST content API this server uses. Manual setup only for now.
3. table_style="plain" (default) — static table. Re-run the report grouped or sorted differently when another view is needed. Good enough for MVP.

Switching to another instance

Only .env changes:

• JIRA_BASE_URL / CONFLUENCE_BASE_URL — set both explicitly if Jira and Confluence are on different hosts (usual for Data Center).
• ATLASSIAN_DEPLOYMENT — leave auto unless the URL doesn't give it away. Data Center means /rest/api/2 and offset-paged search; the client handles the switch, including falling back from the Cloud-only /search/jql.
• JIRA_STORY_POINTS_FIELD — auto-discovery looks for "Story Points" / "Story point estimate"; pin the custom field id if the instance renamed it.
• REFINEMENT_DOR_LABELS / REFINEMENT_DOD_LABELS — the team's actual labels.
• ATLASSIAN_CA_BUNDLE — corporate root CA, for an internal TLS chain. Prefer this over turning ATLASSIAN_VERIFY_SSL off.

Tests

.venv\Scripts\python.exe -m pytest -q

cd frontend && npm test

80 Python tests and 20 Angular tests, no Atlassian instance required — HTTP is mocked with respx, and the metric logic is tested as pure functions over changelog fixtures.

Not built yet

• get_story_comment_summary — a full per-developer comment timeline within a story. get_stale_issues covers the "who has gone quiet" half of it; decide whether the rest is worth having after a few sprints of real use.
• Push alerting (Slack/Teams) for quiet stories. The detection exists; only the delivery channel is missing.
• Postgres sync. Today the dashboard reads Jira live behind a 120 s cache, which is fine at team scale. The rows payloads are already shaped for a table-per-report load if history or cross-team rollups are wanted later.
• Auth on the API. It binds to localhost and assumes whoever reaches it is you. Anything beyond your own machine needs real authentication in front of it.