Skip to main content

Native voice with a local tool

Install @agentium/core, ws, and zod; set OPENAI_API_KEY. The host supplies microphone capture and playback in the negotiated audio format.
This connection example does not play audio. Follow voice sessions and the gateway contract for media integration and confirmed playback history.

Process a completed audio file

VoicePipeline makes one model call after transcription; it does not execute an Agent tool loop. Install openai alongside core for the model adapter.
For a streaming pipeline with Agent tools, use AgentVoiceBrain.

Run a browser task

Install @agentium/browser, @agentium/core, playwright, and openai, then install the browser with npx playwright install chromium.
The browser run owns its browser lifecycle. Native browser actions are not governed by the custom-tool execution policy. Plan mode rejects before launch because native actions have no host effect classifier yet. For closed-choice action planning, configure planner: "jev" and its optional provider dependency; retain a chat/vision model in model. See browser agents.