The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection by creating tools that read data from external APIs, databases, and the file system.\n
Ingestion points: Database query results via the pg library, external API data via fetch, and local file system content via readdir and readFile.\n
Boundary markers: No specific delimiters are used in the templates to separate untrusted external data from the agent's instructions.\n
Capability inventory: The provided code templates enable database queries, outbound network requests, and local file system read access.\n
Sanitization: Employs Zod for input schema validation and a basic string-matching check (startsWith('select')) to limit database tools to read-only operations.\n- [EXTERNAL_DOWNLOADS]: The project setup instructions include downloading standard development dependencies and the Model Context Protocol SDK from the NPM registry.\n
Evidence: Commands to install @modelcontextprotocol/sdk, zod, typescript, @types/node, tsx, and pg.\n- [COMMAND_EXECUTION]: The skill provides standard shell commands for project initialization, dependency installation, and server testing via the command line.\n
Evidence: Use of mkdir, pnpm init, pnpm add, and pipes for testing JSON-RPC methods.

mcp