Video frame API
Search video frames by what the shot looks like
71,471 frames from 838 real launch films, product demos, keynotes and show intros — each one described by a vision model for its craft. Query them over HTTP or hand the whole library to an agent over MCP. No account for the read endpoints, and nothing to install.
MCP7 endpoints · 6 languages · read access is free
Reading the library needs no key at all — search, frames and every MCP tool are open. A key is for the calls whose answer depends on who is asking: your saves, your boards, pinning a frame from an agent.
Authorization: Bearer sk_live_…
What you can actually query
Most video APIs index titles, tags and transcripts — words somebody wrote about the video. This one indexes the picture. Every frame carries a caption, a note on how it is composed, the one thing worth copying from it, and a set of named techniques, all of it searchable:
- Shot type
- no-face, corner-inset, wide presenter, medium presenter, close presenter
- Subject
- presenter, UI, product, terminal, text card, b-roll, chart, logo, crowd
- Surface
- camera, screen recording, motion graphics, 3D render, slide, mixed
- Technique
- named, reusable moves — centred wordmark on black, and a few hundred more
- On-screen text
- verbatim, so a title card is findable by what it actually says
- Measured
- luminance, accent share, face position, face area, cut timing
Endpoints
The full reference, with parameters and examples in 6 languages, is on the documentation page.
Writing for an agent? Skip the HTTP client.
The MCP server exposes 6 read-only tools over the same data, with no account, no key and no client library to keep up to date. One command and Claude Code, Cursor or Codex can search the library itself.
claude mcp add stealshot --transport http https://stealshot.com/api/mcpHow the MCP server works
Questions
What is the STEALSHOT video frame API?
A read-only JSON API over a library of real product launch films, demos, keynotes and show intros, taken apart frame by frame. Every frame has been described by a vision model for its craft — framing, composition, lighting, typography, palette and on-screen text — so you can query for what a shot looks like rather than for a video title.
Is it free, and do I need an API key?
The read endpoints are free and need no key or account. A key is only required for the calls whose answer depends on who is asking: your saved frames and your boards. Those are a Pro feature.
How is this different from a video hosting API?
It does not host, transcode or stream anything. Mux, api.video and Cloudinary put your video on the internet; this searches an existing library of other people's finished work so you can study how it was shot.
Can I use it from an AI agent?
Yes, and that is the better way in. The MCP server exposes 6 read-only tools over the same data with one install command, so Claude Code, Cursor or Codex can search the library without you writing an HTTP client.
What languages are the examples in?
cURL, JavaScript, Python, Go, Ruby, PHP. Every sample is generated from one endpoint definition, so they cannot drift apart.
Is there a rate limit?
There is no published limit and no meter. The read endpoints are database queries over pre-computed descriptions and cost a query rather than a model call. That is a courtesy rather than a guarantee — sustained abuse gets 429.
The frames themselves belong to whoever made the videos; what we license is the description and the index over them. See the acceptable use policy.