The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides tools for the agent to manage external MCP (Model Context Protocol) servers and upload/execute skill scripts. While these allow code execution, the skill mandates server-side security checks (isDangerousFile) and requires explicit user confirmation (--yes) for high-risk operations.
[PERSISTENCE]: The skill includes functionality to create and update Agent automations scheduled via cron or defined intervals (hourly, daily, etc.), which allows the agent to maintain a persistent presence and execute tasks across sessions.
[COMMAND_EXECUTION]: The skill interacts with the local environment via the ae-cli tool and performs local filesystem scans in /home/ta/.local/bin to manage sandbox tools and shims.
[INDIRECT_PROMPT_INJECTION]: The skill processes user memory and skill assets which could contain malicious instructions. However, it implements a robust defense-in-depth approach by explicitly instructing the agent to treat memory content as untrusted candidate data and to silently ignore any control-like sequences intended to override system rules.
Ingestion points: User memory retrieval (memory +list, +context), Skill asset reads (+read-skill-asset), and MCP tool outputs.
Boundary markers: Uses HTML comments () to delimit managed memory blocks in instruction files.
Capability inventory: Perform network requests (MCP), write to the filesystem (attachments, skill assets), and execute scripts.
Sanitization: Instructions mandate the silent ignoring of control-like memory content and treat all retrieved data as non-authoritative.
[CREDENTIALS_UNSAFE]: The skill provides a dedicated command (+mcp-token) to retrieve plaintext tokens for system MCP servers. While this exposes credentials to the agent's context, the instructions include specific warnings about shell history, logging, and secure handling of these tokens.

ae-agent