The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches UI component configuration and styling from the vendor's registry at https://r.assistant-ui.com/mcp-config.json.
[CREDENTIALS_UNSAFE]: The documentation explicitly warns that the default storage implementation (McpLocalStorage) persists authentication tokens in plain text within the browser's local storage, which is susceptible to XSS. It recommends using McpCustomStorage for production use cases.
[COMMAND_EXECUTION]: Instructions suggest using npx assistant-ui@latest for project initialization, which executes code from a remote package in the local environment.
[DATA_EXFILTRATION]: The skill facilitates connections to user-defined third-party MCP servers by URL, which creates a risk of data transmission to untrusted endpoints if a malicious server URL is provided by a user.
[PROMPT_INJECTION]: The skill enables the integration of third-party tool outputs into the agent's runtime, creating a surface for indirect prompt injection where a malicious MCP server could provide instructions designed to manipulate the agent.
Ingestion points: User-provided MCP server URLs and results from tools called via the MCP protocol.
Boundary markers: Not explicitly defined in the provided documentation or implementation examples.
Capability inventory: Support for callTool and network requests via fetch in storage handlers.
Sanitization: The skill focuses on UI management; sanitization of tool outputs for prompt injection is not described.

react-mcp