Your footage, for any AI assistant.

The tools.

• See the library: every synced folder and Frame.io project with ids, the people you’ve named, and any clip’s facts (frame rate, resolution, start timecode, speech, people).
• Find: search Visual (what’s on screen), Spoken (what was said) or by person, inside chosen folders, with filters for length, 4K or HD, frame rate, vertical, shot date, real footage only, and shots already in Selects. Every answer says what couldn’t be searched (a drive that isn’t connected, footage not ready yet). Find shots that look like any clip.
• Check: look at the exact frame at any second, a contact sheet of a clip, and Sift’s shot check: how a stretch moves (locked off, smooth, handheld, shaky), where focus lands, exposure, and sound problems like wind or clipping.
• Organize: work in Selects like an editor (add, remove, trim, reorder, rename, notes), name people, and follow your house rules. Every change is listed in Sift and can be undone.
• Your timeline (with the Sift panel open in Premiere or Resolve): read the open sequence, find alternatives for the shot under the playhead, fill a gap, and place b-roll over an interview or sermon on a new timeline. Your sequence is never changed.
• Deliver: send to Premiere, Resolve or Final Cut, export, gather files, and share by link. Each asks you first, on a card in Sift.

Reading tools are marked read-only. Tools that change Sift (Selects, people’s names) are marked not destructive, and every change is listed and can be undone (undo_change). Anything that leaves the computer or drives another app is two steps: a prepare_* tool (send to editor, export, gather, review link, share folders, b-roll timeline) says exactly what will happen, and run_prepared runs it only after the user clicks Allow in Sift. Results are JSON (structured content) with JPEG thumbnails attached as images. Filenames, notes and transcript excerpts are labelled as data, not instructions.

One click in Sift.

In Sift, open Integrations and click Connect on the Claude or ChatGPT tile. Sift writes only its own entry, keeps the rest of the assistant’s settings and saves a backup first. See Connect to Claude and Connect to ChatGPT.

Set it up by hand.

Any MCP client that can run a local (stdio) server can start Sift’s connector. Sift must be installed and running. On a Mac the command is /Applications/Sift.app/Contents/MacOS/Sift --assistant; on Windows it’s %LOCALAPPDATA%\Programs\Sift\Sift.exe --assistant. In a JSON mcpServers config (Claude desktop and many others):

{
  "mcpServers": {
    "sift": {
      "command": "/Applications/Sift.app/Contents/MacOS/Sift",
      "args": ["--assistant"]
    }
  }
}

In a TOML config (the ChatGPT desktop app’s Codex settings, ~/.codex/config.toml):

[mcp_servers.sift]
command = "/Applications/Sift.app/Contents/MacOS/Sift"
args = ["--assistant"]
startup_timeout_sec = 60

Security and privacy.

The connector only talks to the Sift app on 127.0.0.1; it opens no network port and ignores proxies, and tool inputs can’t choose an address or a route. It can’t delete, move or edit footage. Changes in Sift (Selects, people’s names) are listed and can be undone; sending to an editing app, exporting, gathering files and sharing are prepared first and run only after the user allows them in Sift. Absolute file paths, face identities and face data are never returned; original video never leaves the computer. What the assistant asks for (clip names, folder names, timecodes, quotes, small thumbnails) is processed by the assistant’s maker. More on the privacy page.

When Sift fits.

Sift is built for people with lots of video on their own drives (editors, churches, agencies, documentary and wedding filmmakers) who want to find a shot by describing it, find a line someone said, or find every clip of a person, without uploading footage. It isn’t a cloud asset manager, a frame-comment review tool, or a way to search sound effects; for footage that lives only in a cloud service, search it there.