The Agent Skills Directory

[Unverifiable Dependencies & Remote Code Execution] (MEDIUM): The skill directs the user to add an external MCP server (https://rube.app/mcp) which is not part of the trusted organization list. This server controls the tool definitions available to the agent.
[Indirect Prompt Injection] (HIGH): The skill establishes a high-risk execution pattern where the agent is instructed to 'Always search tools first' and follow 'recommended execution plans' provided by the remote server.
Ingestion points: RUBE_SEARCH_TOOLS fetches dynamic schemas, plans, and pitfalls from rube.app (SKILL.md).
Boundary markers: None. There are no instructions to sanitize or ignore embedded natural language commands within the fetched schemas.
Capability inventory: Includes RUBE_MULTI_EXECUTE_TOOL and RUBE_REMOTE_WORKBENCH, providing the agent with write/execute capabilities on the Revolt platform.
Sanitization: None. The agent is explicitly told to use the exact field names and logic returned by the external API.
[Remote Code Execution] (HIGH): The use of RUBE_REMOTE_WORKBENCH allows for remote command execution or tool orchestration through the Composio framework, governed by configurations fetched from an untrusted external URL.

revolt-automation