The Agent Skills Directory

External Specification Downloads: The skill is designed to clone the UCP protocol specification from an external repository (github.com/Universal-Commerce-Protocol/ucp). While fetching a specification is standard for protocol implementation assistants, it involves downloading third-party content that should be verified for integrity.
Command Execution: To facilitate project setup, the skill executes commands such as git clone and suggests installing several Node.js dependencies (zod, jose, uuid, mcp-handler). The skill explicitly recommends a specific version of the @modelcontextprotocol/sdk to ensure known vulnerabilities are avoided.
Indirect Prompt Injection Surface: The scaffolded code generated by the skill includes a mechanism for capability negotiation that fetches data from a remote platformProfileUrl. This design pattern introduces a surface where the resulting application could process untrusted data from external agents.
Ingestion points: The lib/ucp/negotiation.ts file in the generated scaffold fetches JSON data from a remote URL provided via headers or tool arguments.
Boundary markers: The generated code does not include explicit delimiters or instructions to ignore potential injection markers within the fetched profile content.
Capability inventory: The scaffolded handlers are capable of performing network operations (fetching profiles, processing payments) and managing checkout state.
Sanitization: While the generated code uses zod for structural validation of inputs, it does not include specific sanitization logic to filter natural language instructions that might be embedded in the profile metadata.
Secret Management Best Practices: The skill provides guidance on handling sensitive credentials like JWS signing keys and OAuth secrets. It correctly instructs the user to store these values in environment variables rather than hardcoding them into the project configuration.

ucp