Hivemind Content Studio

A desktop studio for making pictures, video and finished social clips — on your own Mac by default, on a rented or hosted GPU when you ask for one. It is one app rather than a launcher: the same prompt box, the same encrypted library, and the same "where does this run and what does it cost" answer in every studio.

Using the app? → The guide · Troubleshooting Running it from this checkout? → Operations Everything else → docs/

---

What it looks like

A left sidebar in three tiers, and one workspace beside it.

Create — Image, Video, Story, Restore, with a folded Labs group holding Sprite and Lip sync. Each is a canvas with a prompt bar along the bottom and a settings panel down the side. The prompt bar is five things: attach references, load a starter, improve the prompt, choose where it runs, and Generate. The Runs on chip reads out the answer before you press — "This Mac · Z-Image Turbo — free, stays here", or a rented card's name and its hourly price — and defaults to Automatic, which prefers free and local and says why it chose what it chose.

Produce — Planner (say what you want in words; a brain plans it and you confirm before anything is spent), Library (everything you have made, sealed with your key), Productions (the durable runs, with their steps, artifacts and one bounded next action), Inspo (Civitai's gallery, with a button that loads a prompt straight into the studio it belongs in) and Models (install, inspect and remove local weights — the only door to doing so).

Advanced, collapsed — Rented GPUs, Providers, PassBook, Canvas (the node workflow editor), Agents & API, Settings, About.

First launch is two steps at the machine: name the studio and choose a passphrase, then keep the vault recovery key it shows once. There is no compiled-in password, and no server-side copy of yours.

Install and run

From this checkout — the path that works today. Four install commands, the embedded build, two checks and a stack start, listed in full under Developer quick start. Then open the studio at the loopback address the supervisor prints, and follow the guide from the setup card onward.

The packaged app is a signed macOS DMG carrying its own Python, Node and ffmpeg, supervised by the Rust shell in desktop/src-tauri: it reserves a port, starts the sidecars, shows a boot screen with an action per failure, and opens the studio once it is ready. ComfyUI, the Swift MLX engines and rented boxes stay attach-only — the app finds them and never starts or kills them. docs/RELEASE.md is the specification for what that download contains and names exactly what is still unbuilt, so read it before assuming a build exists.

What is inside

The canonical Python package owns every run, asset, provenance, approval, publishing and metrics decision. Everything else is an adapter over it.

Runs are durable: SQLite is authoritative for run, step, event and budget state; a versioned manifest is authoritative for the brief, provider selections, artifacts, provenance, approvals and receipts. Agents ask for capabilities (generate_keyframes, animate_scenes); the router picks a ready provider under the run's privacy, allowlist and budget policy. Paid work uses an HMAC-signed, exact-scope, one-time operator receipt, and rendering never implies publishing.

Prompts, media, workflows and generation parameters stay on the client. Provider secrets live in PassBook, the machine's shared credential store, and never enter the browser or a catalog response.

Documentation

Licensing

Auto Clipper is AGPL-3.0-or-later, so the combined work is AGPL-3.0-or-later. Read THIRD_PARTY_NOTICES.md before distributing. The app's own About page carries the version, the licence, the no-warranty line, the third-party notices and the offer of Corresponding Source.

No credential values belong in this repository; .env.example contains key names only.