Probably developer docs
Build on Probably's AI UGC video production from your own agents and code: the MCP server, the HTTP API with API keys, and signed webhooks.
Probably makes short-form UGC video ads with a crew of AI specialists that research, write, cast, film, edit and review. The same production that runs the studio is open to your own agents and code, so you can drive it from the tools you already use instead of rebuilding it from raw models.
What you get
A harness built on raw video and voice models has to find the route that accepts this week's inputs, retry refusals without paying twice, keep a presenter's face and voice the same across fifty videos, make the footage read as a phone shot it, and catch a wrong product claim before it ships. Probably does that work for every call, on the same credits, spend caps and audit as the studio.
- Outcomes, available now. Whole jobs: make a video from a brief, make variants, make a new version, apply a change, approve a delivery and share it. The crew plans, writes, generates and reviews.
- Stage primitives, starting now. Each step on its own, taking your material and returning typed results you can keep or feed back. Research (
research), hooks checked against your Space's earlier work (brainstorm), reviewed scripts (write_script), the critics' review of your own copy, hooks, video or image (review), storyboards with a reviewed still for each scene (storyboard), reviewed clips of your own scenes (generate_clip) and reviewed recordings of your own lines (voice) are live, metered and capped; swaps and the phone finish are coming.
Three surfaces, one contract
Every surface offers the same tools, with the same permissions, request keys and spend confirmations.
- The MCP server for AI apps: Claude, ChatGPT, Cursor, VS Code, Claude Code, Codex and opencode. See Connect an AI app.
- The HTTP API for your servers, CI and automation tools, called with an API key. See the HTTP API reference.
- Events and webhooks for what happens next: a run started, waiting for a slot, a stage done, waiting on you, delivered or failed. See Events and webhooks.
How a video gets made
Connect
Connect an AI app with OAuth, or create an API key in Settings → Connections. Either acts as you, in one organization and optionally one Space.
Read
List your Spaces, personas, products and videos to find what you work on.
Prepare
Call prepare_video with a brief and a length: 15, 30, 45 or 60 seconds. Nothing is spent: you
get a preview of what it costs and a confirmation.
Confirm
Once the person agrees, confirm_spend starts the work. Confirming again starts nothing more.
Follow
Wait with wait_for_video, read events by cursor, or receive them by webhook until the video is
delivered.
Answer check-ins
A video prepared with mode: "interactive" asks first: before any clip it waits for the person
to look over its storyboard, then to hear its voice. get_video's waitingOnYou names each,
with its scenes or its flagged words, and approve_storyboard, approve_listen, their skip_
tools, keep_word and fix_word answer it.
Start here
- In an AI app or coding agent: Claude Code, Codex or opencode.
- From your own code: curl or TypeScript.
- How keys and connections work: Authentication.
For agents
These docs are also plain Markdown: an index and every page in full. The HTTP API's contract is its OpenAPI document, which the API reference on this site is generated from.