The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill configuration in .mcp.json specifies the use of agentphone-mcp@0.2.0, which is a vendor-provided package managed via the Node.js package runner. This is the intended method for extending the agent's capabilities with the AgentPhone service.\n- [PROMPT_INJECTION]: The skill's primary function involves reading external data such as SMS messages and call transcripts, creating a surface for indirect prompt injection where malicious instructions embedded in incoming communications could influence agent behavior.\n
Ingestion points: Data is brought into the agent context through tools like get_messages, get_call, and stream_transcript in references/api-reference.md.\n
Boundary markers: The skill documentation does not define specific markers or delimiters to separate untrusted telephony content from system instructions.\n
Capability inventory: The agent has access to impactful tools including send_message, make_call, and update_agent, which could be targeted by successful injection attacks.\n
Sanitization: There are no documented mechanisms for filtering or sanitizing the text content received from phone calls or SMS before it is processed by the AI.

agentphone