MCP server for Financial Modeling Prep API - Investment Banking Analyst tools for M&A analysis
All 7 tools are explicitly defined with @mcp.tool() decorators in server.py. Basic structures are present: tool names follow verb_noun conventions (get_*, search_*), descriptions exist for all tools, and input schemas are visible in the decorator parameters. However, critical gaps limit usefulness: (1) Descriptions are extremely brief (10-30 chars), well below the 194-char baseline for production tools, providing minimal guidance on when to select each tool vs. alternatives. (2) Input schemas lack explicit type declarations and JSON Schema structure in the decorator, parameters are inferred from function signatures rather than formally registered. (3) No output schemas are documented; callers cannot predict response structure. (4) Parameter descriptions are missing entirely, e.g., 'ticker' has a brief note but 'period' in get_income_statements lacks any explanation of valid values. (5) No error handling guidance, tools return raw API errors (e.g., {'error': str(e)}) without recovery hints. (6) No pagination despite tools returning lists (limit=4 default suggests potential for large results). (7) The manual tools_registry is defined but appears unused. This server addresses a real domain (financial data APIs) but lacks the robustness expected for agents that need clear selection criteria and error recovery paths.
Fetch latest balance sheets for a company.
Fetch latest cash flow statements for a company.
Fetch company profile, sector, industry, description, etc.
Fetch enterprise value metrics for a company.
Fetch latest income statements for a company.
Fetch financial ratios for a company (e.g., PE ratio, ROE).
Tool descriptions are under 20 characters and provide almost no context for LLM selection. E.g., 'Fetch latest income statements for a company.' (48 chars) vs. production baseline of 194 chars average. No guidance on when to call this vs. similar tools, what the response contains, or when it is useful.
Input parameters lack descriptions entirely. 'ticker' appears in multiple tools with no explanation of format (e.g., 'Stock ticker symbol in uppercase, 1-5 characters, e.g., AAPL'). 'period' in get_income_statements has no note on valid values ('annual' vs. 'quarterly' are inferred from code, not documented). 'limit' defaults to 4 with no explanation of why or valid range.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 40 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 49 | - | v1 |
Search companies by name on NASDAQ exchange.
No documented output schemas. Tools return raw dicts from await fetch_fmp(), but callers have no idea what fields will be present, how they are named, or what types they contain. LLM cannot plan downstream operations or extract required data.
No input schema validation or constraint declaration. The 'period' parameter in get_income_statements is free-form (could be 'annual', 'ANNUAL', 'yearly', etc.) with no enum constraint. 'limit' is unbounded (could be 1 or 10000) with no min/max. LLM will hallucinate invalid values; API will reject them with raw error messages.
Error handling returns raw API errors ('error': str(e)) with no recovery guidance. If FMP API rate-limits or returns 401 (bad key), LLM sees {'error': 'httpx.HTTPError(...)'} and has no idea whether to retry, ask the user, or fix the key.
No pagination or result size limits. Tools accept a 'limit' parameter (default 4), but no guidance on maximum allowed value, total count, or next_cursor. If FMP returns 1000 financial statements, the tool dumps all of them into context.
Tool names are sometimes ambiguous. 'get_ratios' could retrieve dozens of different financial ratios (P/E, PEG, ROE, debt-to-equity, etc.). No name distinguishes this from a hypothetical 'get_valuation_ratios' or 'get_leverage_ratios'.
FMP_API_KEY is stored in environment and used directly, correct pattern for secret injection. However, no documentation in tool descriptions about what happens if the key is missing or invalid (error will be raw API 401).
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All tools are read-only (they fetch data), but this is not declared. An agent could reason that these tools have side effects when they do not.