jira-aie-mcp

An MCP server for the AIE team's Jira instance. Exposes the two things mcp-atlassian can't do natively:

1. Rich Atlassian Document Format (ADF) issue creation/update — info / warning / success / error / note panels, inline status lozenges, date chips, expand/collapse blocks, and task lists. Accepts a Markdown string, a compact block-list DSL, or a pre-built ADF doc.
2. Working-day gantt scheduler — reads Blocks dependencies and Original Estimates from a JQL-scoped set of issues, computes start/end dates on the Norwegian public-holiday calendar, and optionally writes the end dates back to each issue's duedate field.

Plus a semantic link shortcut (jira_link_issues) that picks the right inward/outward direction for Blocks, Relates, Duplicates, Clones, and Causes without forcing the caller to memorise Jira's link model.

mcp-atlassian still handles everything else — search, transitions, comments, watchers, worklogs, etc. This server is a gap-filler, not a replacement.

Install

Run it on demand with uvx — no install step needed; uvx fetches the package from PyPI into an ephemeral environment on first use and caches it for subsequent runs:

uvx jira-aie-mcp

Or install it as a persistent tool:

uv tool install jira-aie-mcp
jira-aie-mcp

Either form starts the server on stdio, which is what MCP clients expect.

Configure (MCP client)

All configuration lives in the MCP client's server definition — no separate config file. Minimum required env vars:

Generate an API token at https://id.atlassian.com/manage-profile/security/api-tokens.

Cowork / Claude Desktop config

{
  "mcpServers": {
    "jira-aie": {
      "command": "uvx",
      "args": ["jira-aie-mcp"],
      "env": {
        "JIRA_BASE_URL": "https://apply-team.atlassian.net",
        "JIRA_EMAIL": "you@apply.no",
        "JIRA_API_TOKEN": "atlassian-api-token-here"
      }
    }
  }
}

To pin a specific version (recommended, so upstream releases don't silently change behaviour):

"args": ["jira-aie-mcp==0.1.2"]

Reload the client after editing the config.

⚠️ If you ran 0.1.0 against a real Jira instance with write_duedate=True, the persisted duedates are inverted — upgrade to ≥0.1.1 and re-run schedule_compute(..., write_duedate=True) once to overwrite them. See CHANGELOG.md for the fix details.

Tools

Block DSL (descriptions)

[
  {"h": 2, "text": "What broke"},
  {"p": "Reranker started returning 500s on ~30% of requests at 14:20 UTC."},
  {"panel": "warning", "content": [
    {"p": "Customer-facing; search relevance degraded but not down."}
  ]},
  {"h": 2, "text": "Timeline"},
  {"ul": [
    "🔴 14:20 UTC — first 500s observed",
    "🟡 14:26 UTC — paged on-call",
    "🟢 14:41 UTC — rolled back cross-encoder to v1"
  ]},
  {"h": 2, "text": "Follow-ups"},
  {"tasks": [
    {"text": "Postmortem", "done": false},
    {"text": "Add 5xx alarm to SLO dashboard", "done": false}
  ]}
]

All block keys (full list):

h      — heading (1-6)                   p      — paragraph
ul     — bullet list                     ol     — numbered list
code   — code block (+ `lang`)           quote  — blockquote
rule   — horizontal rule                 panel  — info|note|success|warning|error
status — inline status lozenge           date   — inline date chip
expand — collapsible block               tasks  — task list with done flags
table  — {headers, rows}                 raw    — pass-through ADF node

Inline specs (inside p / h text or a list item):

{"text": "bold", "marks": ["strong"]}
{"text": "link", "link": "https://..."}
{"mention": "account-id", "text": "@Alice"}
{"emoji": ":rocket:", "text": "🚀"}
{"status": "In progress", "color": "yellow"}
{"date": "2026-04-20"}
{"hard_break": true}

Scheduler

The forward pass follows Blocks dependencies, uses Jira's timetracking.originalEstimate (1w 2d 3h 30m syntax) for durations, and skips weekends and 12 Norwegian public holidays per year (Easter-derived).

schedule_compute(
  jql="project = AIE AND \"Epic Link\" = AIE-5 AND statusCategory != Done",
  start="2026-05-04",
  format="markdown",
  write_duedate=False
)

Returns a JSON plan (project_start, project_end, critical_path, issues[], warnings[]) and, for markdown/mermaid formats, a rendered string.

Rescheduling is just re-running with overrides:

pins   = {"AIE-15": "2026-06-01"}      # force start date
delays = {"AIE-7": "3d"}               # shift start N working days later

Module layout

src/jira_aie_mcp/
  __main__.py     # entry point — starts FastMCP over stdio
  server.py       # MCP tool registrations
  config.py       # env-var loading → Config dataclass
  client.py       # httpx-based Jira REST v3 client + link shortcuts
  adf.py          # block DSL + Markdown → ADF
  holidays.py     # Easter + Norwegian red days + duration parser
  scheduler.py    # graph build + forward pass + critical path
  formats.py      # JSON / Markdown / Mermaid plan output
tests/
  test_scheduler.py   # Blocks-direction contract + diamond end-to-end

Development

Install with the dev extras and run the test suite:

uv pip install -e ".[dev]"
pytest

The test suite is intentionally small and focused on the pieces most likely to regress (the Blocks-direction contract, the working-day forward pass, and the critical path). The REST client and the FastMCP wiring are exercised via the live smoke tests in the team's Jira — if you want those to be unit-level too, mock httpx.Client.post / .put.

License

MIT.