The Agent Skills Directory

[COMMAND_EXECUTION]: The server implementation in scripts/server.py utilizes asyncio.create_subprocess_exec to spawn ffmpeg processes. These are used for audio transcoding (specifically converting MP3 streams to mu-law for Twilio). The command-line arguments are hardcoded for format conversion and do not accept arbitrary user input.
[DATA_EXFILTRATION]: The skill is configured to record and log conversation data, including transcripts and metadata, to a local calls/ directory. Users should be aware that sensitive information discussed during phone calls is stored in plain JSON files on the host system's disk.
[EXTERNAL_DOWNLOADS]: The skill integrates with multiple well-known technology services: Twilio (telephony), Deepgram (transcription), OpenAI (AI reasoning), ElevenLabs (speech synthesis), and Brave (web search). The server scripts connect only to the official API endpoints for these services.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. Because the system transcribes user speech and appends the resulting text directly to the AI's conversation history without sanitization or strict boundary markers, a malicious caller could attempt to override the agent's persona or manipulate its use of tools like web search.
Ingestion points: Incoming audio stream from Twilio calls, processed via Deepgram or the OpenAI Realtime API.
Boundary markers: Transcripts are interpolated directly into the message history without delimiters or instructions to ignore embedded commands.
Capability inventory: The agent can perform web searches using the Brave API and has the ability to save data to the local filesystem.
Sanitization: Transcripts are processed in their raw form by the language model.

phone-agent