Picking a voice normally means auditioning a list. Asking for 'warm, confident female narrator' and getting three candidates is the difference between a two-minute detour and a twenty-minute one. The other design worth noting is stop_video_agent_session halting at a checkpoint and preserving partial results — most generation APIs give you cancel-and-lose-everything. Be deliberate about delete_video: the tool description says permanent and irreversible, and an agent has no way to undo it.
HeyGen's endpoint, one of the largest documented surfaces here: 72 tools across 17 categories. The Video Agent group is the headline — describe what you want and it handles scripting, avatar choice and scene composition itself.
- Video Agent: create from a prompt, list styles, poll the session, send follow-up messages, stop at a checkpoint with partial results kept
- Videos: list with folder and title filters, create from an avatar or an arbitrary image with a script or pre-recorded audio, get status and URLs, delete permanently
- Templates: list, read the variable schema with defaults, generate by substituting text, image, video, audio and character variables
- Voices: list by engine, language and gender; design_voice returns up to three matches for a description like 'warm, confident female narrator'; clone_voice from an audio file
- And the rest across translation, AI clipping, batches, avatars, assets and lipsync
OAuth 2.0 against HeyGen's authorisation server, with device flow and client credentials also supported. MCP usage draws on your existing subscription credits; API-key usage bills the API plans separately.
Paste a URL, then authorize — add the endpoint to your client, then approve the OAuth consent screen