The agent answers questions about our architecture but misses decisions documented in Confluence — no error, just confident incomplete answers. Why?
The silent failure of permission-trimmed search: Confluence page restrictions. Restricted pages are excluded from the integration search index entirely — not shown as locked, not flagged, simply absent — so the agent synthesizes an answer from the visible subset and presents it as complete. The engineering fix: run all agent access through a dedicated service account, then deliberately grant that account access to the engineering spaces that constitute your knowledge corpus; treat spaces it cannot see as non-existent, because for the agent they are. Periodically audit by asking the agent to enumerate the spaces and pages it can access and diffing that against reality.
Hosted official server vs the community self-hosted bridge?
The hosted mcp.atlassian.com endpoint is OAuth-secured, Rovo-backed, and permission-trimmed automatically — the enterprise-safe default. The community bridge (sooperset/mcp-atlassian) offers deeper raw API control, JQL passthrough, and Data Center support, at the cost of owning API token hygiene yourself. Cloud teams start hosted; Data Center holdouts have exactly one option.
Can the agent create and transition Jira issues?
Yes — create, update, comment, and transition through workflow states, scoped to the authenticated account's permissions. Ground the agent on project keys and workflow state names first; ungrounded transitions are the leading cause of issues parked in phantom statuses.
How does it compare to Linear MCP?
Depth versus breadth. Linear MCP is a sharper instrument for a single opinionated tracker; Atlassian MCP spans Jira plus Confluence — the ticket AND the spec — which is the compound workflow that actually matters in enterprises. Choose by where your organization's context lives, not by feature checklists.