MU MCP Server provides 14 well-intentioned tools for semantic code analysis. Tool naming follows verb_noun convention (mu_search, mu_expand, mu_read) which is appropriate for internal DSL. Descriptions are present for all tools and generally clear (median ~120 chars, within baseline range). Parameter schemas are fully specified with types (usize, u8, bool, string, arrays) and descriptions. However, several critical gaps reduce overall score: (1) Output schemas are NOT documented, the tools return markdown strings with no formal field definitions, forcing LLMs to parse unstructured text; (2) No error handling guidance, tools lack actionable recovery instructions or error classification; (3) Parameter defaults are underspecified, many accept optional params without explicit default behavior documented; (4) No pagination/limits guidance beyond inline comments; (5) Tool composition issues, mu_review combines mu_diff + mu_impact + mu_audit but the integration is not formally documented; (6) Security considerations for WRITE tools (mu_configure, mu_enrich, mu_bootstrap) lack permission gates or audit trail declarations; (7) STDIO-only transport hard-caps protocol readiness at 50, preventing remote accessibility. Despite these gaps, the server demonstrates solid foundational quality: parameter lenient deserializers show care for client tolerance, tool descriptions explain WHAT and WHEN to use each tool, and the semantic code analysis domain is well-scoped.

• Document output schemas for all tools as formal JSON Schema structures. E.g., mu_search should return {results: [{node_id: string, confidence: number, file: string, line: number}], total: number, limited: boolean}. This enables LLMs to reliably extract and chain results.
• Add error handling guidance to every tool description. E.g., 'If a node_id is invalid, mu_read returns an error with suggestions for similar node names. Use mu_search to discover valid IDs.'
• Make default parameter values explicit in descriptions. Change 'maximum number of results (default: 10)' to 'maximum number of results; returns up to 10 by default'.
• Expand mu_sus description to 'find_suspicious_code' or rename with full name to avoid ambiguity.
• For mu_configure, provide an enum or schema documenting valid correction keys: service_classifications, domain_concepts, add_priority_nodes, remove_priority_nodes, test_handling. Include example JSON structure.
• Add pagination support to mu_search, mu_expand, and mu_audit. Accept offset/limit parameters and return {results: [...], total: number, hasMore: boolean} so LLMs can paginate large result sets.
• Document mu_review composition formally: state which fields come from mu_diff (changed_nodes), mu_impact (downstream_effects), and mu_audit (violations). Include example response structure.
• For WRITE tools (mu_bootstrap, mu_configure, mu_enrich), add a 'Permissions' section in the description declaring required scopes (e.g., 'write:analysis_db', 'read:source_code'). Note whether confirmation is required before execution.