Corbell MCP server has 4 tools with adequate naming and moderate descriptions, but suffers from incomplete parameter documentation, missing output schemas, and insufficient error handling guidance. All tools are read-only, which is positive for safety, but the lack of structured output definitions and parameter constraints significantly reduces usability for LLMs. Schema documentation is minimal, parameters exist but lack type constraints and validation rules. Error handling returns generic error strings rather than actionable recovery guidance. This places the server in the C/D range (fair-to-poor) on the calibration scale.

• Document output schema for each tool. E.g. graph_query returns {service: {id, name, language, type, repo, tags}, dependencies: [{target_id, kind}], methods: [{name, class_name, file_path, line_start}]} with field types and counts.
• Add minimum/maximum constraints to numeric parameters: top_k_services (e.g. 1-50, default 10), top_k (e.g. 1-100, default 10), include_dependencies and include_methods as booleans with clear descriptions.
• Revise error handling to categorize failures: 'Service not found' → suggest running `corbell graph build` first; 'Workspace not configured' → include setup instructions with environment variable hint (CORBELL_WORKSPACE).
• Add pagination to code_search and list_services: include limit (default 20, max 100), offset/cursor, and total_count in responses. Document that large result sets are capped and batches available.
• Clarify tool selection in descriptions: 'Use graph_query for service metadata and direct dependencies; use get_architecture_context to auto-discover relevant services for a feature description.'
• Add natural-language identifier support hints to descriptions: 'service_id can be a service name or system ID; if unsure, call list_services first.'
• Include chaining IDs in all responses: if get_architecture_context returns service names, also return service_ids so downstream graph_query calls don't require another lookup.