aqm857886159/Nomi

Open-source AI video workbench. Bring any model or your local ComfyUI, and let Claude Code / Codex / Cursor direct it over MCP — storyboard, references, generation, editable first cut on a real timeline. Local-first: projects, prompts, and keys stay on your machine. No account, no telemetry.

Nomi – Open‑source, local‑first AI video studio

What it is – Nomi is a desktop application (macOS & Windows) that lets you create short AI‑generated videos by talking to an on‑screen agent. The agent breaks your description into shots, generates key‑frame images, turns each shot into video, and assembles everything on a timeline. All generation is done through the AI models you connect (cloud APIs or local ComfyUI), so you pay only the provider’s price and keep full control of keys and data.

Why it matters – Commercial AI‑video tools often hide the model behind a subscription and markup the inference cost. Nomi gives a “pro‑grade” workflow while staying model‑agnostic and local‑first: you choose the endpoint, the model, and the cost; the app itself is free and AGPL‑3.0.

Core features (as described in the README)

Conversational agent – type a natural‑language prompt; the agent reads the canvas, plans shots, generates keyframes, renders video, and places everything on the timeline.
Storyboard view – each shot appears as a row with its prompt, first frame, selected model, and duration. Reference cards (character or location) can be locked to keep a consistent appearance across shots.
3‑D director – pose characters, move a virtual camera, and use your phone as a viewfinder; captured frames are fed back into the canvas instantly.
Per‑shot model selection – any supported model (Seedance, Kling, Wan, Hailuo, Nano Banana, GPT‑Image, etc.) can be assigned to an individual shot, letting you mix and match providers.
One‑click model integration – the Let AI connect it for me button runs Claude Code or Codex to look up a new model’s endpoint, parameters, and ID, test it, and add it to the UI.
MCP server – Nomi ships a local “Model‑Control‑Protocol” server with 24 tools, enabling external agents (Claude Code, Codex, Cursor, etc.) to drive generation, arrangement, and editing programmatically.
Project as a folder – a Nomi project is just a directory containing the canvas, storyboard, timeline, and generated assets; no cloud account is required.
Privacy‑first – by default nothing is sent off‑device; anonymous diagnostics are optional.

How it works (high‑level flow)

Connect a model – add an API key for a provider (APIMart, Kie.ai, OpenAI‑compatible relay, local ComfyUI, etc.) or let the built‑in AI fetch the config for you.
Describe your film – type a natural‑language request in the agent panel or start from an existing storyboard.
Agent planning – the agent reads the canvas, splits the story into shots, selects a model for each, and generates keyframes.
Video generation – each keyframe is sent to the chosen model, which returns a short video clip; the clips are placed on the timeline.
Edit & export – lock reference cards, adjust durations, reorder clips, then export the final MP4.

Getting started (quick‑start steps from the README)

Download the appropriate installer (macOS arm64/intel DMG or Windows x64 EXE) from the Releases page.
Launch the app (macOS may require clearing the quarantine flag; Windows may need to bypass SmartScreen).
Connect a model – paste a provider key or add a custom endpoint.
Enter a prompt – tell the agent what you want to make.
Review/adjust the generated shots on the canvas, lock any reference cards, and arrange the timeline.
Export the finished video as MP4.

Technical notes

Built with Electron (main process) and a React + Vite + Tailwind front‑end.
Requires Node 20+ and pnpm for development; pnpm dev runs the app locally.
The repository includes a skills/ folder (Skill Pack v2) that defines the agent’s capabilities.
Optional research tooling can read Douyin/Xiaohongshu/Bilibili APIs using a TIKHUB_API_KEY env var.

License & community

License: AGPL‑3.0‑only for current releases (earlier releases remain Apache‑2.0). Commercial custom work must stay AGPL‑compatible.
Community: GitHub Discussions, Issues, and a WeChat user group. Sponsors (e.g., Gway) are listed but do not affect the open‑source nature.
Contributing: No CLA required; contributions accepted under the same AGPL‑3.0‑only license.

Bottom line: Nomi is a genuine, open‑source AI video creation tool that puts the choice of model and cost in the user’s hands while offering a conversational, shot‑by‑shot workflow suitable for creators who want a desktop‑first, privacy‑preserving solution.