Multi-agent MCP framework for Cursor IDE with semantic routing, persona management, and context enrichment
Server has 8 tools with schemas and descriptions present, but quality is inconsistent. Tool naming follows verb_noun convention (clear, actionable). Descriptions range from adequate (route_and_load: 156 chars) to sparse (clear_session_cache: 79 chars). Critical gaps: (1) No input validation guidance in parameter descriptions, e.g., 'query' param lacks format/length constraints; (2) Output schemas completely undocumented, no return type specs for any tool; (3) Error handling absent, no guidance on retryability, user-fixable vs fatal errors; (4) Parameters like 'protocol_version' and 'current_persona' lack clarity on when/why to use them; (5) Composition issues, route_and_load and get_agent_context appear to do similar work (routing + context loading) but distinction is unclear. Per-tool average: 52/100. Baseline comparison: avg tool description 194 chars (this server ~110 chars), 100% of A+ tools document return types (this server: 0%).
Administrative cache reset. Not needed for persona switches; affects shared caches.
Extract text from an image file using Vision AI. Supports JPG, PNG, TIFF, BMP, WebP, HEIC formats with optional preprocessing.
Extract text from a PDF document using Vision AI. Converts PDF pages to images and processes them with Vision API. Requires poppler-utils system package.
Load enriched context for a specific agent (system prompt, rules, skills, implants). Used after routing to get full agent bundle.
Log user interaction with agent response for analytics and history tracking.
Read conversation history from workspace memory store.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. All 8 tools lack return type specifications.
Parameter descriptions lack validation constraints. 'query' param has no length/format guidance; 'limit' and 'offset' in read_history lack min/max bounds; 'dpi' in extract_text_from_pdf lacks range.
No error handling guidance. Tools do not indicate which errors are retryable, user-fixable, or fatal. No recovery suggestions (e.g., 'If agent not found, call list_agents()' for get_agent_context).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Route user query to appropriate agent and load enriched context (system prompt, rules, skills, implants). Supports both protocol v1 and v2.
Write repository summary and context to workspace memory for future reference.
Unclear tool composition. route_and_load and get_agent_context both load context; distinction is vague. Parameter 'force_reload' and 'refresh_persona_context' suggest overlapping concerns. Agents may call both unnecessarily.
Sparse descriptions for administrative tools. clear_session_cache (79 chars) and log_interaction (60 chars) lack context on when/why to invoke them. Descriptions below 100 chars risk LLM misuse.