GitHub Release+npx Promethee MCP with local browser sign-in and bounded task/project tools.
promethee-mcp demonstrates solid tool design with complete schemas, clear descriptions, and proper input validation via Zod. All 6 tools have descriptions (avg ~80 chars), structured input schemas with type constraints, and output types defined. Tool names follow verb_noun convention (list_, get_, create_). Pagination is implemented with cursor+limit. Idempotency keys on write operations prevent duplicate side effects. Main gaps: descriptions are functional but brief (under 100 chars), lacking WHEN/WHY guidance for LLM selection; no explicit error recovery guidance in descriptions; output schemas documented in contracts but not surfaced in tool definitions themselves; no per-parameter descriptions visible in the tool registration layer (only in Zod schemas).
Check whether this local MCP process is connected and return the browser login URL when needed.
Create one bounded project for the authenticated user with an idempotency key.
Create one bounded task for the authenticated user with an idempotency key.
Get one approved task record owned by the authenticated user.
List the authenticated user's approved project fields with bounded pagination.
List the authenticated user's approved task fields with bounded filters and pagination.
Tool descriptions lack WHEN/WHY context for LLM selection. E.g., 'List the authenticated user's approved task fields' does not explain when to call this vs get_task, or what 'approved' means to the agent.
Parameter descriptions are embedded in Zod schemas (src/contracts/tool-inputs.ts) but not visible in the tool registration layer (src/mcp/create-server.ts). LLMs see only parameter names and types, not the validation rules or format hints (e.g., 'idempotency key' or 'cursor for pagination').
No explicit error recovery guidance in tool descriptions. E.g., create_project with invalid clientRequestId returns a validation error, but the description does not hint at the regex pattern (^[A-Za-z0-9_-]{16,64}$) or suggest how to fix it.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Output schemas (ListTasksResult, GetTaskResult, etc.) are defined in src/contracts/tool-results.ts but not documented in tool descriptions. LLMs cannot see what fields to expect (e.g., 'observedAt', 'freshness', 'nextCursor') without inspecting the source.
Tool descriptions do not clarify scope/permission requirements. E.g., 'approved task fields' and 'authenticated user' are vague, the description should state 'Requires tasks:read scope' to help agents understand authorization boundaries.