Framework-RAI exposes 5 tools with basic schemas and descriptions, but falls short of production quality. All tools have descriptions (10-60 chars, below the 50-200 char baseline for LLM optimization) and input schemas are present but minimal. Parameters lack depth: most have no type constraints, enums, or validation guidance. No output schemas documented. Error handling is absent, no recovery guidance, categorization, or actionable messages. The server uses stateful session management (initialize → session_id → tools/call), which is deprecated in the current MCP spec. Tool naming is verb-forward (scan, generate, analyze, get, update) but descriptions are too terse to guide LLM selection. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk profiles (4 READ_ONLY, 1 WRITE).

• Expand tool descriptions to 50-200 characters. Example: 'scanProject' → 'Scan a project directory for AI/ML components (models, training scripts, data pipelines). Returns a list of detected components with file paths and risk flags. Call this first to understand the project scope before detailed analysis.'
• Document output schemas for all tools. Specify what fields each tool returns (e.g., scanProject returns {components: [{file, type, riskLevel}], totalCount}). This enables LLMs to chain tools and extract data correctly.
• Add enums and constraints to parameters. Replace free-form 'type' with enum: ['checklist', 'model_card', 'risk_file']. Add validation rules in descriptions: 'modelType: one of (pytorch, tensorflow, sklearn, custom) or omit for auto-detection.'
• Implement error handling with recovery guidance. Return structured errors: {code, message, suggestion}. Example: 'Project path not found at /invalid/path. Verify the path exists and is readable. Available projects: /home/user/project-a, /home/user/project-b'.
• Add tool annotations to the tools/list response. Include readOnlyHint: true for scanProject, generateSuggestions, analyzeModel, getDocumentation. Include destructiveHint: true for updateDocumentation. This signals to agents which tools are safe to retry.