The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides tools like mcp__relaycast__agent_add and mcp__relaycast__command_invoke which allow the agent to spawn external CLI processes (e.g., aider, goose, claude) and execute arbitrary registered commands.
[CREDENTIALS_UNSAFE]: The tool mcp__relaycast__workspace_set_key is used to programmatically set and manage workspace API keys, which presents a risk of sensitive credential exposure or unauthorized configuration changes.
[PROMPT_INJECTION]: The skill's core functionality of inter-agent messaging creates a significant surface for indirect prompt injection.
Ingestion points: Tools such as mcp__relaycast__dm_get, mcp__relaycast__message_get, and mcp__relaycast__inbox_check in SKILL.md ingest content from other agents.
Boundary markers: None; messages are processed as raw strings, and there are no instructions provided to the agent to treat this data as untrusted or to ignore embedded commands.
Capability inventory: The skill has access to powerful tools including process spawning (mcp__relaycast__agent_add), command execution (mcp__relaycast__command_invoke), and file uploads (mcp__relaycast__file_upload).
Sanitization: No content sanitization or validation mechanisms are described for incoming messages.
[DATA_EXFILTRATION]: The inclusion of webhook tools (mcp__relaycast__webhook_create and mcp__relaycast__webhook_trigger) allows for the creation of arbitrary network egress points which could be exploited to exfiltrate sensitive data.

using-agent-relay