The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents an experimental 'Code Mode' (enabled via experimental_codeMode) that allows an LLM to generate JavaScript which is then executed in a secure V8 isolate using the secure-exec package. While mitigated by the isolate, this pattern involves executing untrusted code generated at runtime.
[COMMAND_EXECUTION]: The setup and troubleshooting sections instruct the user or agent to execute shell commands, including package installations (pnpm add, npm install) and CLI utility invocations (npx nuxt add mcp, npx @modelcontextprotocol/inspector, curl).
[EXTERNAL_DOWNLOADS]: The skill facilitates the download and installation of multiple Node.js packages from the NPM registry, including @nuxtjs/mcp-toolkit, zod, evlog, and secure-exec. These are associated with the Nuxt ecosystem and official MCP tooling.
[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for building tools and resources that ingest data from external sources, such as third-party APIs (weather.ts) and local project files (read-file.ts, readme.ts). These ingestion points represent a potential surface for indirect prompt injection if the processed content contains malicious instructions designed to influence the agent's behavior.
Ingestion points: Tools fetching from https://api.weather.com/v1/current and reading files via node:fs/promises.
Boundary markers: The examples provided do not demonstrate the use of explicit delimiters or instructions for the LLM to ignore embedded commands in the tool output.
Capability inventory: The framework supports network operations ($fetch), file system access (readFile), and dynamic code execution (secure-exec).
Sanitization: Input validation is performed using Zod, but the examples do not illustrate sanitization or filtering of the content returned to the LLM context.

manage-mcp