Why Swiftbird exists

Making a documentary with AI agents already works. Swiftbird's builder had made films that way before Swiftbird existed: a script researched and written by an agent, a beat sheet of more than a hundred shots, 3D scenes built in Blender by agents working in parallel lanes, graphics in Remotion, narration, and a final edit in Premiere. Whole films were made like that, shot by shot.

The trouble wasn't the agents. It was everything around them.

The work was scattered

A film made this way lives in dozens of files: a script, a beat sheet, lane handoffs, judge records, proof stills, render folders, thumbnails, an editorial project. The agents knew where everything was. The person directing them had to reconstruct it every time: which shot is this, what did I ask for last time, is this still the version I approved?

Corrections were the hardest part. "Move the camera a bit closer and hold on the lantern" is a paragraph an agent has to interpret. Say it three times, across three sessions, and you're managing wording rather than directing a film.

One desk, not a new engine

The idea was simple: keep the workflow exactly as it is, and put a proper desk in front of it.

• One Chat that drives the agent you already use, Claude Code or Codex, with its own skills and sign-in.
• Pages that show the real files, read in place: every shot with its narration and stills, the scene in 3D, the thumbnail versions, the calendar, the upload.
• Corrections that carry their context. A note on a shot, a selected sentence in the script, or a camera path drawn in the scene editor goes back to the agent as an exact, versioned request, not a paraphrase.

Swiftbird isn't a new production engine, and it doesn't pretend to be one. It doesn't render, it doesn't have its own AI, and it doesn't move your files. It centralises the work you're already doing.

Honest by design

Tools like this are tempting to oversell, so Swiftbird is deliberately strict about what things mean:

• an agent's "done" is not your approval;
• a planned date is not an upload;
• a scene-editor proposal is not a change to Blender until the agent builds it;
• a file that's missing is shown as missing.

Those rules come straight from real production. They're what stops a long film from quietly drifting away from what you actually approved.

Why local

Documentary work is heavy and private: gigabytes of scenes, unreleased scripts, licensed footage. There's no reason for Swiftbird to upload it, so it doesn't; only what you send your agent goes to its provider. Swiftbird's cloud holds your sign-in and, once paid plans launch, a minimal billing history. Everything else stays on your machine.

Where it is now

Swiftbird runs every day on its builder's Windows PC. It isn't downloadable yet: the installer, updates and paid plans are still being finished, and the changelog says plainly what has and hasn't been proven. If you make documentaries with AI agents, or want to, join the Windows waitlist.