SYSTEM Cited by 2 sources

Atlassian MCP server¶

What it is¶

The Atlassian MCP server (also called the Atlassian Rovo MCP server) is Atlassian's hosted, remote MCP server at https://mcp.atlassian.com/v2/mcp. One server spans the Teamwork Collection — systems/jira, Confluence, Loom, Goals/Projects — so one agent session can cross products, with the resulting links landing in the Teamwork Graph. Reported scale: >15M tool calls every weekday and ~2M MAU, up 8× since March 2026 (Source: sources/2026-10-28-atlassian-jiras-depth-on-demand-designing-an-mcp-server-around-jobs).

Tool-surface design (v1 → v2)¶

v1 (MVP): 15 Jira + 13 Confluence + a few shared tools, near-1:1 with REST endpoints; 32 definitions (~60 KB JSON, ~15k tokens) loaded up front; creating an issue took four calls; terse descriptions.
v2 (job-first): 52 Jira tools that call 100+ distinct APIs, chosen from user requests, usage rankings (reads and writes ranked separately), and dogfooding — instead of auto-generating tools for Jira's ~725 REST operations. Canonical patterns/tool-surface-minimization instance.
Server-side orchestration: e.g. manageJiraSprint resolves the tenant's story-point custom field, updates estimates, adds issues via the Agile API, and starts the sprint in one call; getJiraBoardSprintData resolves board + active sprint by name and returns progress slices.
Schema tactics: enum action parameters; names instead of IDs; descriptions that say when to use a tool and point to neighbours; errors with the faulting field, allowedValues, and a repairHint; JQL passthrough; view = compact/evidence/full.
Workflow-preserving writes: closing an issue = listJiraIssueTransitions then transitionJiraIssue, so Jira validators and conditions still run.

Context loading¶

Six core tools are always loaded; the other 46 are deferred, found through discover, and run through executeRead, executeWrite, or executeDestructive, so each call's risk tier is explicit (concepts/progressive-capability-disclosure). ?tools=all returns a flat catalog for MCP gateways. Reported: up to 25% fewer tokens than v1; full up-front loading ≈ 2.2× the cumulative input tokens of the deferred design (concepts/context-window-as-token-budget).

Authorization and governance¶

Tool groups by intent: read_jira (27), write_jira (15), search_jira (1) on by default; delete_jira (3) and manage_jira (6) off by default, each behind its own admin-enabled scope (concepts/least-privileged-access).
Calls run with the signed-in user's permissions over OAuth 2.1 (patterns/on-behalf-of-agent-authorization).
Atlassian Guard Data Security Policies can block MCP access per project/space or per classification (Confidential, PII), stricter than browser access (concepts/governed-agent-data-access).
In agentic pipelines, each stage gets scoped MCP tool permissions (Source: sources/2026-08-28-atlassian-agentic-automation-in-practice-putting-standard-engineering-work-on-autopilot).

Evaluation¶

Outcomes are judged by final Jira state, not agent self-report. October 2026 internal 84-case eval: 17 new tasks enabled; 57/62 vs 54/62 on shared cases; failed tool calls per run 0.20 → 0.09. Failing cases are kept as jobs-to-be-done evals to steer improvements.

Seen in¶

sources/2026-10-28-atlassian-jiras-depth-on-demand-designing-an-mcp-server-around-jobs — primary subject: Jira tool redesign, deferred loading, scope groups, DSPs, eval numbers.
sources/2026-08-28-atlassian-agentic-automation-in-practice-putting-standard-engineering-work-on-autopilot — the Atlassian MCP tools used by Rovo Dev in the Dispatcher → Coding Agent → Closer loop, with per-stage scoped tool permissions.