Production-quality TypeScript starter for the Anthropic Claude SDK — built for regulated financial institutions
This is a Claude SDK starter, not an MCP server. It implements a custom tool registry pattern for the Anthropic Claude API, not the Model Context Protocol. Four tools are defined with basic schemas and descriptions, but lack depth. Tool names are action-oriented (calculator, web_search, file_reader), descriptions are present but generic (10-50 chars, below the 50-200 char baseline for LLM optimization). Schemas show required fields but lack comprehensive parameter documentation. No error handling guidance, no output schema documentation, no tool annotations. The file_reader tool has a critical security concern: path validation is mentioned but not visible in the provided code. Overall, this is a functional but minimal implementation suitable for a starter template, not production-grade tooling.
Evaluate a math expression
Read file contents with path safety validation
[MCP/filesystem] List the contents of a directory
Search the web for information
Descriptions are too short (10-50 chars) to guide LLM tool selection. Baseline is 50-200 chars. E.g., 'Evaluate a math expression' lacks context on when to use calculator vs other tools, what it returns, or error cases.
No output schema documentation. LLMs cannot plan downstream calls or extract results without knowing what fields are returned. E.g., does calculator return {result: number} or {value: string, error?: string}?
Parameter descriptions are missing or minimal. 'expression' and 'query' lack format guidance, constraints, or examples. Baseline is 72 chars per param description; these are ~10 chars.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 38 | <=2025-11-25 | v2 |
file_reader claims 'path safety validation' in description but validation logic is not visible in provided code. Either the validation is incomplete or the description is misleading. Path traversal attacks are critical security risks.
No error handling guidance. Tools do not document what errors are retryable, what the LLM should do on failure, or how to recover. E.g., if web_search times out, should the agent retry or ask the user?