MCP server for semantic search and Supabase integration with AI-powered research and analysis tools
Server has 9 tools with basic schemas and descriptions, but significant quality gaps. All tools have descriptions (10-100 chars, mostly adequate), and input schemas are present with type definitions. However, naming lacks action verbs (planner, researcher, analyzer, critic, synthesizer, factchecker are role names, not verb_noun patterns). Parameter descriptions are minimal, most are single-sentence. Output schemas are undocumented; tools return generic {content: [{type: 'text', text: ...}]} with no structured fields for chaining. Error handling is minimal (try/catch with error message only, no recovery guidance). No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Composition is weak: agent tools (planner, researcher, etc.) are thin wrappers around an LLM API with no real integration; they don't compose with semantic search or query tools. No pagination, no result limits documented, no field naming consistency across tools.
The analysis request or data to analyze
The content to evaluate or critique
SQL query to execute (only SELECT queries are allowed)
The statements or claims to verify
The ID of the company to get details for
The planning request or question to address
The research question or information to investigate
Agent tools (planner, researcher, analyzer, critic, synthesizer, factchecker) use role names instead of action verbs. LLMs cannot infer intent from 'planner' alone, should be 'plan_research', 'generate_plan', or similar verb_noun pattern. This violates the 90% baseline of A+ tools starting with action verbs.
No output schemas documented. All tools return generic {content: [{type: 'text', text: ...}]} with no structured fields. LLMs cannot extract IDs, references, or metadata for downstream chaining. Violates pattern:tool and mxe:include-chaining-ids.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 42 | 2026-07-28+ | v2 |
Search query to find relevant companies
The information to synthesize or integrate
Error handling is minimal. Tools catch errors and return {content: [...], isError: true} with only the error message. No recovery guidance, no categorization (retryable vs fatal), no actionable next steps. Violates pattern:recovery-guide.
Agent tools (planner, researcher, etc.) are thin wrappers around an external LLM API with hardcoded system prompts. They do not integrate with the semantic search or query tools in the same server. No composition or chaining between tool groups. Violates pattern:tool-chain.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All tools are marked READ_ONLY in metadata but no annotations in schema. Missing idempotent hints for query tools. Violates current spec (2026-07-28) tool annotations pattern.