Browser-backed MCP server wrapping upstream mcp-atlassian with Playwright SSO auth
Single tool 'atlassian_login' has moderate naming (verb-noun form is present but generic for an auth action) and a reasonable description, but lacks critical schema detail and parameter descriptions. The input schema is visible in the provided source and shows an enum constraint for 'target' and a string 'url' parameter, but parameter descriptions are minimal. There is NO documented output schema, making it impossible for LLMs to know what the tool returns. The tool performs authentication (a stateful, write-level operation) but does not declare error recovery guidance, dry-run support, or clear failure modes. Since this is the only tool and exhibits significant gaps in schema documentation and output specification, the overall definition quality is poor.
Authenticate with Jira or Confluence using browser-based SSO
No output schema documented. LLMs cannot determine what the tool returns (success response structure, auth token format, session ID, error fields, etc.). This violates the critical pattern requirement that all tools must document output schemas.
Parameter 'url' has no description. LLMs cannot infer that this is optional, what format is expected, or when to provide it vs. relying on environment variables.
Parameter 'target' has minimal description ('The Atlassian service to authenticate with (jira or confluence)'). Does not explain WHY there are two separate targets, WHEN to pick each, or whether both can be authenticated in one session.
Tool description does not explain error modes or recovery steps. What happens if the browser cannot be opened? What if SSO fails? What if the user cancels login? LLMs need actionable error guidance.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | <=2025-11-25 | v2 |
No mention of idempotency. Does calling atlassian_login twice with the same target create duplicate sessions, or is it safe to retry? Agents need this information.
Description states tool authenticates 'using browser-based SSO', but does not explain that this is NOT an in-process browser (which would block). The code comment says 'the server never opens a browser itself to avoid hanging', but this critical behavior is not surfaced in the tool description for LLMs.