# Nomi: a local-first AI video workbench that your coding agent can drive over MCP

> Nomi is an open-source Electron app for storyboard, generation and timeline editing that runs against your own model keys or a local ComfyUI. The interesting part is the 25 MCP tools that let Claude Code, Codex or Cursor operate it as a generation backend.

**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.

- Repository: https://github.com/aqm857886159/Nomi
- Website: https://nomiaqm.com
- Stars: 537 · Forks: 121
- Language: TypeScript
- License: AGPL-3.0
- Published: 2026-09-10 · Updated: 2026-09-10 · Language: en
- Canonical page: https://hysenlabs.com/projects/aqm857886159-nomi

## The cost problem Nomi is built around

Generation is where AI video budgets actually go. Nomi's premise, stated in its README, is that you should be able to mix cheap and expensive generation sources freely rather than committing to one platform's pricing. The README names relay wholesale prices, time-limited platform promotions, free image credits bundled with an agent membership, free sources such as ModelScope, and a local ComfyUI as combinable options.

The workflow it recommends on top of that is a draft, reference, finish ladder. Cheap or free models produce storyboard sketches, pose frames and reference videos, which you can regenerate without much thought. The takes that survive get fed as references into a higher-quality model, so the expensive credits are spent only on the final pass. That is a normal production discipline, but the app is structured to make it cheap to follow rather than something you do by hand across five browser tabs.

The audience is narrower than the tagline suggests. This is for people generating video at volume, or companies that want an internal AI video pipeline, not for someone who wants to type a prompt and get a finished clip with no further decisions.

## What the local-first claim actually covers

Projects, prompts and keys stay on the machine. There is no account and no telemetry, per the README and the package description. When Nomi calls an external model API, only the inputs required to complete that task go to the provider you configured. That is a meaningful distinction from browser-based generation tools, where your reference images and intermediate takes live on someone else's storage by default.

The trade-off is that nothing is managed for you. Provider credentials, model selection, and the question of which endpoint is cheapest at any given moment are yours to handle. The README acknowledges this directly, calling the integration burden the one big pain point, which is why it offers a second path: hand a document to your coding agent and let it walk you through wiring up a relay, DeepSeek, or a local ComfyUI.

## MCP as the structural difference

Nomi exposes 25 MCP tools so that Codex, Claude Code or Cursor can drive it for generation, orchestration and editing. The README's argument is that an online platform cannot offer this because it monetises compute, while a local open-source app has no reason to refuse being used as a backend.

The practical consequence: if your agent membership bundles image or video credits, an agent that can call Nomi can spend those credits on generation tasks inside a real editing surface. You write the intent in the agent, the agent calls the tools, and the results land in a project you can cut on a timeline.

The repository layout supports the claim rather than just asserting it. There is a .mcp.json at the top level, alongside .claude/, .codex/ and .agents/ directories, plus a docs/integrate-with-your-agent-en.md document that the README points to. The integration document is described as covering four paths, each with a success signal: custom or relay providers, local ComfyUI, MCP driven by an agent, and skill import.

## Installing Nomi and connecting a first model

Releases are published for macOS arm64, macOS x64 and Windows x64. Linux, Windows arm64 and a macOS universal installer are not currently published, so a Linux workstation is out of scope for the prebuilt route.

On macOS, download the matching DMG and drag Nomi.app to Applications. The build is not Apple Developer ID signed or notarized, so the first launch is blocked. The README's procedure is to right-click the app in Applications, choose Open, then confirm Open. If that fails, open System Settings, go to Privacy and Security, find the Nomi message and click Open Anyway.

If macOS reports the app as damaged, the README's instruction is to first confirm the installer came from an official Nomi link, then clear the quarantine attribute:

```bash
xattr -dr com.apple.quarantine "/Applications/Nomi.app"
```

The README explicitly says not to disable Gatekeeper globally. On Windows the installer has no Authenticode signature, so the SmartScreen prompt requires More info, then Run anyway.

After launch, the quick start is a two-step loop. First connect a model: pick a curated provider and enter one API key, or add your own OpenAI-compatible, Responses-compatible or Anthropic-compatible endpoint. Second, write the intent, starting from a story or a single shot. APIMart and Kie.ai are wired in as the two cores, with around ten more ready-to-use providers including ModelScope, Volcengine, Runway, fal, Replicate, MiniMax and ElevenLabs.

A local ComfyUI is treated as a provider like any other. Nomi converts the normal Save workflow format, so downloaded workflows import, and it diffs the graph against /object_info to report which custom nodes and model files are missing before you run it. If you would rather not work this out yourself, clone the repository and hand docs/integrate-with-your-agent-en.md to your agent with a description of what you want connected. Building from source requires Node 22.19.0 or newer and pnpm 10.8.1, per package.json.

## Where Nomi gets in your way

Updates are manual. The README states that updating means downloading the matching DMG and replacing the app. There is no in-app updater described, and no documented rollback path if a new release breaks your provider configuration. If you run Nomi across several machines, that is a repeated manual step each time.

The unsigned macOS build is not a one-time inconvenience. Every fresh install on a new machine hits the same Gatekeeper sequence, and the quarantine command is a workaround rather than a fix. For a team deploying to non-technical editors, that friction is real.

ComfyUI integration depends on graph inspection, not on running the workflow. Nomi reports missing custom nodes and model files by diffing against /object_info, which catches structural gaps. It cannot tell you whether a workflow will produce acceptable output, and a graph that passes the diff can still fail at execution for reasons the check does not model.

The provider list is a starting point, not a guarantee. The README describes a continuously growing integration-certification ledger with 66 certified entries at the time of writing. Anything outside that ledger is your problem to debug, which is exactly the situation the agent-assistance document exists to handle.

## How Nomi differs from an integrated generation platform

The obvious alternative is a hosted AI video service that bundles model access, storage and editing in one subscription. Those services handle procurement, so you never think about which endpoint is cheapest, and they handle updates, so you never clear a quarantine flag. The difference in approach is that they own the compute relationship. Your footage and intermediate takes live in their storage, and the model menu is whatever they have negotiated.

Nomi inverts that. You own the provider relationships and the files, and the app is a workbench on top. The cost argument follows from the inversion: because you can route a storyboard pass through a free source and a final pass through a paid one, the per-unit cost is a function of your routing decisions rather than a published price list. The README claims per-unit cost drops sharply under this workflow. Treat that as the project's stated rationale, not a measured result.

A second alternative is assembling the same pipeline yourself from a ComfyUI instance, a script directory and a video editor. That gives you complete control and no Electron app to trust. It also means building the storyboard surface, the reference management and the agent tool definitions yourself, which is most of what Nomi is.

## Licence, maintenance and upgrade cost

Nomi is licensed AGPL-3.0-only, and the repository carries a CLA.md alongside the LICENSE file. AGPL-3.0 is a strong copyleft licence with a network-use clause. If you modify Nomi and let users interact with it over a network, the licence's terms about offering corresponding source are likely to apply. The README also mentions AGPL-compliant custom development and deployment through the maintainer, which suggests the project anticipates commercial users who need terms beyond the default. This is not legal advice; read the licence text and talk to counsel if you plan to redistribute or host a modified version.

Maintenance looks current. The repository is not archived, and the last push was on 2026-09-10. Releases are frequent: v0.21.0 on 2026-08-26, v0.20.1 on 2026-08-21 and v0.20.0 on 2026-08-17. That cadence means the upgrade cost is mostly the manual DMG replacement plus whatever provider or workflow configuration a release changes. The CHANGELOG.md at the repository root is where that would be documented, and checking it before replacing an install is the cheapest way to avoid a surprise.

## Conclusion

Adopt Nomi if you already pay for an agent membership or run ComfyUI locally and want a timeline editor that consumes those sources without moving your footage off disk. Do not adopt it if you need Linux builds, signed macOS installers, or a managed cloud service that handles model procurement for you. Before committing, verify two things against the current release: that your provider or ComfyUI workflow imports cleanly, and that the AGPL-3.0-only terms fit how you intend to distribute anything you build on top of it.

## FAQ

### Does Nomi run on Linux?

No prebuilt installer is published for Linux. The README lists macOS arm64, macOS x64 and Windows x64 as the supported release targets, and states that Linux, Windows arm64 and macOS universal installers are not currently published.

### Does Nomi require an account or send telemetry?

The README states there is no account and no telemetry, and that projects, prompts and keys stay on your machine. When Nomi calls an external model API, only the inputs required to complete the task are sent to the provider you configured.

### Can Claude Code or Cursor control Nomi?

Yes. Nomi exposes 25 MCP tools so that Codex, Claude Code or Cursor can drive it for generation, orchestration and editing. The repository includes a .mcp.json file and a document at docs/integrate-with-your-agent-en.md describing the integration paths.

### How do I connect a local ComfyUI to Nomi?

Nomi treats a local ComfyUI as a provider like any other. It converts the normal Save workflow format so downloaded workflows import, and diffs the graph against /object_info to report which custom nodes and model files are missing before you run it.

## Sources

- [aqm857886159/Nomi on GitHub](https://github.com/aqm857886159/Nomi)
- [License: AGPL-3.0](https://github.com/aqm857886159/Nomi/blob/main/LICENSE)
- [Project website](https://nomiaqm.com)
- [README](https://github.com/aqm857886159/Nomi/blob/main/README.md)
- [Releases](https://github.com/aqm857886159/Nomi/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/aqm857886159-nomi