HITE

Vibe coding for video editing. Describe an edit in plain language and a real timeline cuts it.

Website · Docs · Architecture · Contributing

There is no black box between the prompt and the result. The model emits typed edit commands, those commands reduce into a versioned edit decision list, and that list compiles into the scene graph that renders. You can undo any of it, or ignore the prompt entirely and drag the clips yourself.

The film in the middle of that clip is HITE's own export, the mp4 its worker rendered and uploaded to its exports bucket, watermark included. The title cards around it, and the editor footage in the full 53-second version, were assembled afterwards in ffmpeg.

That timeline, read out of the database rather than remembered: 21 clips over 25.00s cut to a 120 BPM grid, 20 transitions across 7 kinds, look-a24, 53 effect applications across 8 keys, two overlays, and a text-lower-third-basic title. The footage is the synthetic sample this repo ships (pnpm sample:take).

HITE is self-hosted

There is no hosted editor. You clone this repo, point it at your own Supabase, bring your own model key, and run it. Your footage never leaves your machine and nobody pays for your renders but you.

https://www.tryhite.xyz runs this code as a landing page. It has no Supabase project behind it, so /app there answers a 503 naming the variables it wants. That is the correct behaviour for a deployment meant only to serve the landing, and it is why the site's buttons say Source rather than Open the editor.

Status

• The async backend runs end to end. Queue, analyze, render, MP4. Covered by an integration test that spawns the real worker, drives real ffmpeg and a real headless Chromium, uploads to real object storage, and ffprobes the result. Four tests, all passing against a local stack.
• The catalog over-promises. The registry advertises 76 effects, looks and transitions; 46 actually change the picture. The gap is measured and gated, so the model is never offered the other 30 (lib/ai/tools/renderableEntry.ts).
• The browser UI has no automated coverage in the default gate. Everything under it does. The Playwright suite in qa/ walks into /app, but pnpm test does not run it, so assume the timeline UI has edges the tests do not see.
• Every provider ships untested. Eight are wired (Google, OpenAI, Anthropic, Groq, xAI, DeepSeek, OpenRouter, self-hosted OpenAI-compatible). The badges come from public/providers/verified.json, which only scripts/verify-provider.ts writes.

Quickstart

Verified on Windows 11 with Node 24.16.0, pnpm 10.18.0 and Docker Desktop 4.67.0.

You need Node 20+, pnpm, Docker (for the local Supabase), and an AI provider key, which you paste into the app's settings panel rather than into .env. Gemini's free tier is enough to try it: https://aistudio.google.com/apikey.

git clone https://github.com/Imdevsup/hite.git && cd hite
pnpm install

The Supabase CLI will not work straight after install. Its postinstall is blocked by pnpm.onlyBuiltDependencies, so node_modules/supabase/ has no bin/. Fix it from inside its own package directory, because the script reads package.json relative to the working directory and dies on the repo root's:

cd node_modules/supabase && node scripts/postinstall.js && cd ../..

Then bring up the database and the app:

pnpm supabase start        # applies supabase/migrations/*.sql automatically
cp .env.example .env.local # fill in the three SUPABASE_* values from `pnpm supabase status`
pnpm dev                   # http://localhost:3000
pnpm worker                # SECOND TERMINAL, see below

This project's Supabase ports are not the defaults: API 54421, DB 54422, Studio 54423, Mailpit 54424. Anything you read elsewhere saying 54321 is wrong for this repo.

pnpm worker is a second process and the app is broken without it. It drains the Postgres job queue, which is every analyze and every export. Without it running, uploads and the editor work fine and every job sits queued forever.

Open http://localhost:3000/app. The settings sheet opens on first visit and asks for your provider key. Nothing is stored server-side and no account is created: middleware.ts mints an anonymous Supabase session, so anonymous sign-ins must be enabled on the project. That is already set for local in supabase/config.toml; on a hosted project it is Authentication, then Sign In / Providers, then Anonymous. With it off, /app answers a 503 saying exactly that.

How it works

One spine, and nothing bypasses it:

EditCommand[]  ->  reduceBatch  ->  Edl.2  ->  edlToRenderIR  ->  HiteRoot
lib/edl/commands   lib/edl/reducer  lib/edl/schema  lib/render/compile  lib/remotion

• The model and your hands go through the same door. The planner and the manual UI emit the same command union through the same reducer, which is why they share one undo stack.
• HiteRoot is the only compositor. The browser <Player> and the worker's renderMedia run the same tree over the same IR, so the preview is the export. ffmpeg does decode, encode, probe and analysis, never compositing.
• Time is integer ticks at 30000/sec. ticksToFrame is called only by the compiler.
• Nothing is fabricated. Missing analysis is reported as missing. A look step that cannot resolve is dropped with a diagnostic rather than guessed at.

The AI layer is bring-your-own-key. There is no deployer key pool and nothing falls back to one, so a request with no key is a 402 pointing at settings. The key arrives per request in a header and is never stored. lib/ai/providers/registry.ts imports no provider SDK, which is what keeps the browser bundle at zero; factory.ts is the only module that does.

Analysis and render go through a Postgres job queue (claim_job, reap_stale_jobs) drained by worker/index.ts. It reaps jobs abandoned by a dead worker, heartbeats the ones it holds, enforces a per-job deadline, and fences every terminal write on its claim.

The long version is in docs/ARCHITECTURE.md and on the docs site: the command union, the reducer's invariants, the tool router, and what the renderer cannot draw yet.

Tests

pnpm typecheck      # tsc --noEmit
pnpm exec vitest run
pnpm lint

With no environment configured that is 111 files / 1577 tests passing, plus 8 files / 65 tests skipped. The skipped ones are the integration suites, which need a database.

To run those, start the local stack and export SUPABASE_LOCAL_URL, SUPABASE_LOCAL_ANON_KEY and SUPABASE_LOCAL_SERVICE_KEY from pnpm supabase status. On Windows also point REMOTION_BROWSER_EXECUTABLE at an installed Chrome, or the render tests hang for four minutes before failing for the wrong reason.

They are not mocks. They run the real worker against real Postgres and real object storage, and they prove row-level security at runtime rather than asserting it.

Deploying

You do not need this section to use HITE. Running it on your own machine is the supported path.

The Next.js app deploys to any Node host. The worker does not: it is a long-lived process that renders video, so it needs a container or a VM with the same repo, NEXT_PUBLIC_SUPABASE_URL, SUPABASE_SERVICE_ROLE_KEY, ffmpeg (vendored) and a Chromium. Deploy the app without it and every analyze and export job queues forever.

A deployment with no Supabase configured is a landing page. middleware.ts answers /app/* with a clean 503 naming the variables it wants.

Contributing

CONTRIBUTING.md has the setup, the gates and the conventions. The short version: match the surrounding code, ground claims in the source rather than in memory, and never report a number the code does not actually produce.

Issues and pull requests are welcome. If you are looking for somewhere to start, the effects catalog is the obvious one: 30 of the 76 registry entries have no renderer, and each one is a self-contained piece of work with a clear test.

License

HITE's own source is MIT.

It bundles ffmpeg binaries via ffmpeg-static, which are GPL/LGPL, and GSAP, under its own licence. If you ship HITE commercially those terms are yours to satisfy. The third-party notice in LICENSE names them.