nomi

Nomi is an open-source, local-first desktop application for AI video production, distributed for macOS (Apple Silicon and Intel) and Windows x64. It aims to cover the whole pipeline in one workbench: writing intent, building an editable storyboard and canvas plan, generating images or video with configured models, arranging results on a real timeline, and exporting MP4. Bring your own models Nomi does not proxy or resell inference. You connect providers with your own keys: two curated providers (APIMart and Kie.ai) are wired in, alongside roughly ten more ready-to-use integrations such as ModelScope, Volcengine, Runway, fal, Replicate, MiniMax and ElevenLabs. Any OpenAI-compatible, Anthropic, Responses or relay endpoint can be added by pasting a URL and a key. A local ComfyUI instance is treated as a provider too: Nomi converts the standard "Save" workflow format so downloaded workflows import, and diffs the graph against /object_info to report missing custom nodes and model files before a run. The README mentions a continuously growing integration-certification ledger (currently 66 certified entries). Agent-driven workflow The project exposes 24 MCP tools so that Codex, Claude Code or Cursor can drive Nomi for generation, orchestration and editing. A dedicated document, "Let your AI connect Nomi for you", is written to be handed to a coding agent, covering four paths: custom/relay providers, local ComfyUI, MCP-driven operation, and skill import. The repository also ships a Skill Pack v2 under skills/, documented in docs/skill-pack-format.md. Local-first and privacy Projects, prompts, generated takes, workflows and provider configuration are stored on the user's machine. No account and no telemetry are required; when an external model API is called, only the inputs needed for that task are sent to the configured provider. The README discloses that one curated provider link carries a referral code, while users always pay providers directly with their own keys. Cost-oriented workflow The stated motivation is that generation is the dominant cost in AI video. Nomi encourages mixing cheap or free generation sources (relay pricing, promotions, bundled agent credits, free sources, local ComfyUI) with a "draft to reference to finish" pattern: cheap models for storyboard sketches, pose frames and reference videos, then feeding selected results as references into a higher-quality model for the final step. Development and community Building from source requires Node.js 20+ and pnpm; the repo layout separates the Electron main process (local runtime, storage, model calls) from the React + Vite + Tailwind workbench. Checks include pnpm run test, pnpm run typecheck and pnpm run gates. A research script (scripts/research/tikhub-search.mjs) searches Chinese social platforms for creator commentary and reads its credential only from the TIKHUB_API_KEY environment variable. Releases are AGPL-3.0-only, with historical Apache-2.0 releases keeping their original license; contributors do not sign a CLA. Feedback channels include GitHub Discussions and Issues, plus WeChat groups for Chinese-speaking users. Paid AGPL-compliant custom development, integration, deployment and training are offered for teams.