Hickeyfield

A free, open-source desktop studio for AI video and image generation. Bring your own API keys and pay providers directly at cost — there is no subscription, no credits, no plan gating, and nothing to pay us.

A hickey is the speck of dirt that prints as a haloed blemish — the flaw that proves a thing went through a real press rather than a renderer.

Hickeyfield is an independent open-source project. It is not affiliated with, endorsed by, or sponsored by Higgsfield, Inc.

Status: early, and honest about it. Hickeyfield generates: you add a provider key, pick a use case, attach media if the job needs it, and get real video and images back, saved to a folder you choose. What is verified end to end is the fal route on macOS — upload, submit, poll, download, play. Windows builds in CI on every commit but has not been exercised by hand. The surfaces beyond the generator (library, audio, studios) are not built yet.

Why

The best AI media tools are harnesses: they wrap third-party models in presets and a good interface. The interface is the product, and the interface is reproducible. What isn't reproducible is subsidised bulk model pricing — which is exactly what a subscription buys you, and exactly what you don't need if you already hold your own API keys.

So: same features, same workflow, your keys, at cost.

What that costs

On flagship video you may pay slightly more per generation than a bulk subscriber, because large platforms buy in volume and subsidise. On Veo Fast/Lite, Grok direct, Gemini Omni, Seedream, Recraft, upscaling and every self-hosted path, you pay considerably less. And you never pay for a month you didn't use.

Hickeyfield always shows the real USD cost before you submit, computed from live provider price feeds rather than an opaque credit integer.

What it does today

The generator is organised by use case rather than by model, and each one offers only the models that can actually do that job — so a model that cannot take a video never appears under Edit Video, instead of failing after you have attached one and pressed Generate.

• New Video · text to video
• Animate Image · a still becomes a shot
• Edit Video · change something in footage you already have
• New Image · text to image

Everything a request depends on is checked against the provider's own published schema before anything is sent: which inputs the endpoint accepts, which values it enumerates, and what shape the result will be. A setting the endpoint does not have is reported rather than silently dropped, and a value it will not accept is refused before you are charged rather than after.

Prompts can be expanded through a filmmaking corpus — camera, lens, light, motion — either deterministically from a preset, or through a local model via Ollama. Nothing is sent anywhere for this unless you ask for it.

Architecture

Tauri v2 — a Rust core with a React UI in the system webview.

crates/hickeyfield-core/   provider adapters, routing, job engine, presets, costs
                        (no Tauri dependency, so tests run in seconds)
src-tauri/              the desktop shell: commands, events, window lifecycle
ui/                     React + Vite frontend
scripts/                signing, verification and provenance tooling
provenance/             one-way hashes used by the copy-provenance lint

Being native rather than a web app is load-bearing, not incidental:

• No CORS, so every provider is reachable directly, including the several that refuse browser requests outright.
• No proxy, and therefore no server that could hold or leak a key.
• Keys live in the OS keychain — macOS Keychain, Windows Credential Manager — and never cross into the webview. The UI can ask whether a key is set; it can never read one.
• Jobs run in the Rust core, so they survive the window closing and reattach on relaunch.
• Export uses native FFmpeg with hardware encoding, so there is no length or resolution ceiling.

Provenance

Two CI lints protect the line between reimplementing a workflow and copying someone's work:

• No third-party media hosts may appear anywhere in the repo. Every preview and sample is generated by us.
• No verbatim third-party product copy in strings we ship. Short functional labels are unavoidable; anything longer is written by us. This is checked against one-way hashes, so no third-party text is ever stored here.

The model catalogue is seeded from higgsfield-ai/cli's MODELS.md, which is MIT licensed and published by its authors as a machine-readable specification. See NOTICE.

License

AGPL-3.0-or-later. See LICENSE, and NOTICE for bundled third-party components.