Your agent can write the code.
It has never seen a good launch video.

Search the frame library by what is actually IN and TRUE OF a shot. This is the main way in. Every frame has been described by a vision model, so plain language works: 'presenter against a dark backdrop with the UI floating beside them', 'code on a light background with huge negative space', 'title card that opens on the wordmark'. Combine free text with the filters to narrow by craft.

query · technique · subject · surface · shot_type · tone · brand · launch_only · opening_only · limit

Given a technique or a plain-language move, gather the EVIDENCE for it across different videos: one best example per brand, plus how widespread it is. This is the tool for 'is this a real convention or did one video do it once', and for building an argument rather than a single reference. Always prefer this over search_frames when the question is about a pattern rather than a picture.

technique · query · limit

List the technique vocabulary the library actually uses, with how many frames and how many distinct videos use each. Call this before search_frames when you want to filter precisely, or to answer 'what moves are common in launch videos'. A technique used by many DIFFERENT videos is a convention; one used many times inside a single video is that video's habit, so both counts are returned.

contains · min_videos · limit

The full anatomy of one video: its measurements plus a shot-by-shot walk through its described frames in time order. Use this after search_frames when a single reference is worth studying properly, or to answer 'how is this video actually built'. Prefer it over many search calls on the same title.

video_id · title · max_frames

Compare how two or more brands cut their videos: pace, how much of the frame is a face, how dark they shoot, and their most-used techniques. Use for 'how does Linear's launch style differ from OpenAI's'.

brands

What is actually in the library right now: counts, the brands with the most indexed videos, and how much of it has been described. Call this first if you need to know whether a question is answerable, and to avoid claiming coverage the corpus does not have.

no arguments

What is an MCP server?

Model Context Protocol is a standard way for an AI agent to reach data it was not trained on. STEALSHOT's server lets your agent search a library of real launch videos frame by frame and cite what it finds.

Do I need an API key or an account?

Neither. It is a public, read-only server. One command and your agent can use it.

Which clients work?

Claude Code, Claude Desktop, Cursor, Codex, VS Code, Windsurf, and anything else that speaks Streamable HTTP.

Can it change anything in my project?

No. Every tool is annotated readOnlyHint in the protocol itself, so your client can verify that rather than take our word for it. There is no write path, so there is nothing to approve and nothing to revoke.

How were the frames described?

Every frame is passed to a vision model in batches, in time order, with the running order in the prompt, so a shot is described as a move rather than as an inventory of objects. It records framing, composition, lighting, on-screen text, palette and a named technique. Face position, face area, luminance and cut timing are measured separately with a detector, not guessed.

What is actually in the library?

Product launch films, feature announcements, demos and screen recordings, keynotes, and podcast and show intros. Both the well-known launches and the ones that only did well on X. It is growing, and anybody can add a URL.

How fast is it?

Under a second for most calls. The tools are database queries over pre-computed descriptions, not model calls, so there is no inference in the request path.