The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements an architecture where interactive UI widgets receive data from an external MCP server and can communicate back to the agent using the App class. Specifically, the app.sendMessage and app.updateModelContext methods allow data originating from the server to inject messages or context updates into the conversation, potentially influencing the agent's next steps.
Ingestion points: Data enters the widget via the ontoolresult event handler, which receives and parses JSON content from the MCP server (e.g., in SKILL.md and references/widget-templates.md).
Boundary markers: The provided templates do not include specific boundary markers or warnings to the agent to ignore instructions embedded in the server data.
Capability inventory: The skill utilizes tools capable of sending messages to the agent, updating model context, and calling other server-side tools.
Sanitization: While the skill details robust UI-level sanitization (CSP and iframes), it does not explicitly address the sanitization of data that might be interpreted by the LLM as instructions.
[COMMAND_EXECUTION]: The documentation provides examples of shell commands for testing and installation, such as npm install, npx mcp-remote, and standard JSON-RPC loop testing using cat and jq. These are standard developer operations and do not represent malicious intent.
[EXTERNAL_DOWNLOADS]: The skill references official documentation for the Model Context Protocol and mentions IP ranges for Google and Anthropic for rate-limiting configurations. These references target well-known and trusted infrastructure services.

build-mcp-app