Docs

The Agent-Native SDK · 10.04

Media, Vision & Transcription

Provider-pluggable image, audio, speech, and video generation, plus vision, transcription, and document extraction feeding the same content-addressed asset store.

MediaVision

A tiny registry, not a hardcoded provider

MediaService is deliberately thin: providers register the modalities they can produce, and agentis.media.generate dispatches to whichever provider actually handles the requested modality. Nothing is hardcoded — the default OpenAI-compatible image provider is registered from environment configuration at bootstrap, and adding audio, video, or a completely home-grown generation backend is one call to register(), no core platform code touched.

Everything generated becomes the same kind of reference

Every generated result is persisted straight to the content-addressed asset store and comes back as an artifact:<id> reference — the exact same shape a screenshot or an uploaded document uses. That uniformity is what lets chat and every channel render a generated image or an extracted document with zero modality-specific rendering code; the surface only ever has to know how to render one kind of reference, regardless of what actually produced it.

Continue