This MCP server demonstrates strong naming conventions, excellent parameter documentation via Pydantic Field annotations, and comprehensive schema information visible in the source. All 17 tools follow verb_noun naming (list_*, get_*, run_*, generate_*, fetch_*, save_*, apply_*). Each tool has a clear, substantive description (50-150 chars). Parameters are richly annotated with descriptions, examples, regex patterns, and type constraints (FastMCP generates inputSchema from Pydantic Field metadata). Output schemas are inferred from return type hints (Envelope[dict] pattern). However, three issues prevent a higher score: (1) Output schemas are not explicitly documented in tool definitions, they are inferred from return annotations, making them less visible to clients; (2) Error handling guidance is minimal, tools return errors via the Envelope wrapper but lack actionable recovery hints; (3) The 'save_strategy' tool accepts a JSON body as a string parameter, which is non-standard for agent composition (most tools should accept structured parameters, not opaque JSON strings). Tool definitions are directly visible in src/alpha_forge_mcp/server.py with explicit @mcp.tool() decorators, so all tools are scored based on actual code, not inference.

• Add explicit outputSchema fields to each tool definition. Instead of relying on return type hints, declare the response structure: {type: 'object', properties: {ok: {type: 'boolean'}, data: {type: 'object', properties: {...}}, error: {type: 'string'}}}.
• Refactor save_strategy to accept structured parameters (strategy_id, version, timeframe, indicators, rules, etc.) instead of a json_body string. This enables agents to compose strategies programmatically without string manipulation.
• Enhance error responses with actionable recovery hints. Example: if forge binary is missing, return {ok: false, error: 'forge binary not found', nextSteps: 'Install AlphaForge CLI from https://alforge.io/install', retryable: false}.
• Document timeout and duration expectations for long-running tools. Add fields to descriptions: 'Expected duration: 30-300 seconds. If no progress update for 60 seconds, the operation may be hung.'
• Explicitly mark read-only tools in descriptions with '(idempotent, safe to retry)' to clarify which tools agents can freely invoke without side effects.
• Add a confirmation_required parameter to destructive operations (run_optimize with save=true persists data). Offer a dry_run mode to preview results before committing.
• Document parameter relationships (e.g., start_date and end_date are inclusive; if both omitted, defaults to cached range). Add examples in descriptions showing typical parameter combinations.
• Return paginated results for list_* tools. Add limit and offset (or cursor) parameters to cap results at 50 items and enable iteration over large result sets.