velorn is a production layer wrapped around a ComfyUI you already run

Editing, captions and export work without ComfyUI

The dependency split is stated in two sentences near the top and everything else follows from it. Editing, captions, export, project management and the MCP editorial tools work without ComfyUI at all. All current generation features require a locally running ComfyUI instance. That is a narrower and more useful promise than a feature list, because it tells you the editor is not a thin front end: the timeline, the caption renderer and the export queue are the product, and the generation panel is where the dependency lives. The stated scope is one project-based application bringing planning, generation, asset management, timeline editing, captions, effects and export together, rather than a set of tools that each talk to a different backend. The framing sentence at the top of the page is the same idea in one line: the agent generates media, builds the timeline and mixes the audio, live, over MCP. Which means the agent path and the generation path share a dependency, and if ComfyUI is down the agent can still edit. The editor surface is listed in enough detail to judge it. There is a project asset browser and a multi-track video and audio timeline, with trimming, moving, snapping, overlap replacement behavior and transitions on clips. Text, shape, title, solid-color, adjustment-layer, keyframe and visual effect tools sit alongside inspector controls, and proxy and cache tools exist for smoother playback. Captions are generated from the edited timeline audio rather than from a separate script, so transcription is timeline-aware, and the styling covers font, color, outline, background, shadow and animation, with presets that can be saved for reuse and a live preview with play and scrub controls plus safe-zone overlays. Export carries practical render presets, hardware-accelerated options where available, numbered PNG image sequence output, queue controls and project-aware output settings. A separate Stock tab searches Pexels and imports photos or video straight into the current project, and the Pexels API key is optional, added in Settings once you have one.

It calls itself a production layer, explicitly not a replacement

There is a paragraph that reads like a positioning statement written to pre-empt the obvious comparison. For generation, the project is not a replacement for ComfyUI; it is the production layer around ComfyUI. The four verbs are the whole argument: plan the work, send jobs to ComfyUI, collect the outputs, and finish the edit. Each of those is a stage ComfyUI does not do, and the reason the distinction matters is that the failure mode people expect is a second, worse node editor. What you get instead is an application that already knows what a project is, what a shot is and what an export preset is, and hands the parts it is not better at to something else. Three ways in are offered. Built-in local and cloud workflows. Your own ComfyUI API workflow JSON. Or the bundled Velorn Bridge, which exists so a graph opened in ComfyUI can be sent back into the right Velorn panel. That third one is the interesting piece, because it closes the loop in the direction people actually work: build a graph in the tool that is good at graphs, then push the result somewhere that knows about timelines. The page also states what the project is for in six bullets, and every one of them is about producing something editable rather than a finished file: music videos assembled from lyrics, timing, characters, keyframes and shot edits; creator-style and small-business ads built on editable shot plans; curated local and cloud workflows run from a single Generate workspace; custom ComfyUI workflows for image, video, keyframe and music-video work inside the app; generated clips edited with tracks, transitions, effects, captions and proxy tools; and generated media, prompts, workflow outputs and timelines kept together inside one project.

Generation runs local, cloud and your own ComfyUI graphs, on loopback only

Three kinds of workflow sit behind the Generate tab, and the split matters more than the model names do. Local workflows cover image, video, image-edit, audio and utility jobs. Cloud workflows run on partner nodes where they are available, and the ones named are Nano Banana 2, GPT Image 2, Seedance and Kling. Custom Image and Custom Video workflows are for people who want Velorn to run their own ComfyUI API graphs, with JSON import as the manual path for anyone who prefers to export workflows from ComfyUI themselves. Underneath all three is a check step that looks for missing nodes, models, credentials and configuration before a job leaves the app, which is the difference between a workflow that fails on the first run and one that fails with a message naming the missing piece. Browsing is a three-way split as well, a Featured view, My Workflows and Templates, with Local and Cloud filters, and imported community workflows appear in Featured next to the built-ins. The Templates tab is the widest surface in the product: it browses the official ComfyUI template catalog of more than five hundred entries, shows size and popularity information for each, and launches any of them into the embedded ComfyUI tab. The integration section is specific in a way that tells you exactly what is supported. Velorn talks to a local ComfyUI server and can also help launch it. The default endpoint is a fixed local address on the conventional ComfyUI port, http://127.0.0.1:8188, and a custom port is available in settings. Launching is handled per platform: a Windows launcher for a configured ComfyUI start script, and a macOS launcher for a configured application bundle. Optional auto-start, stop-on-quit and restart behavior are all there, which means the lifecycle is something you configure once rather than manage by hand. The embedded ComfyUI tab exists for opening and editing graphs in place, including signing into a ComfyUI account inside that tab and showing a credit balance when one is available, so a cloud-node workflow is something you can set up without leaving the app. The constraint is the line to remember: only localhost and loopback ComfyUI endpoints are supported in the desktop app. There is no remote ComfyUI, which is a sensible boundary for something that shells out to arbitrary graphs, and also the thing to check before you plan a setup with ComfyUI on another machine.

Four guided workflows, one of them admitted beta

The Create tab is where the guided paths live, built on an engine the project calls Director Mode, and there are four of them with a clear target audience each. Music Video Creation turns a song, lyric timing, characters, references and a director script into keyframes, video shots and an editable timeline. The UGC Creator builds creator-style social ads with hooks, dialogue, product demos, try-ons and testimonials, with editable shot-by-shot output. The Business Ad Creator is offer-first, aimed at local businesses, ecommerce products, events, services and small teams. And Short Film Creation is described as an experimental script-to-scene coverage workflow that is still very beta and may have rough edges. That last label is unusually honest for a feature list, and it is the one to plan around: the three named ad and video workflows are the product surface, the film one is an experiment. What the three have in common is that they all end in something editable rather than a finished file, since a shot plan you can still adjust is the difference between a generator and an editor.

The music video creator names its model routes explicitly

This is the workflow with the most detail in the README and the details are the reason to read it. Song import and lyric timing come first, and lyrics can arrive either through speech recognition or by pasting them and aligning them into a subtitle file. Cast setup accepts existing character sheets, so an existing cast does not have to be retyped. Then it is per shot: keyframe prompts, reference images, prompt copying, prompt editing, image replacement, and rerunning a single shot. Both keyframes and video have built-in routes and a custom path, and the custom path is where the control lives, with optional injected keyframe image, prompt, seed, width, height, frame rate, duration and audio. The built-in routes are named, which is rarer than it sounds: Qwen Image Edit and Nano Banana 2 on the keyframe side, LTX 2.3 Music and WAN 2.2 on the video side, with custom keyframe workflows going through Velorn endpoint nodes instead. The last step is timeline assembly from the generated shot assets, which is the point where a pile of clips becomes something with structure. Pinning a seed and a duration per shot is what makes a rerun comparable to the previous one.

A hundred MCP tools, and a test script per feature

The agent story is a local MCP server with more than a hundred tools, named as working with Codex, Claude Code, Cursor-compatible tools and other MCP clients. Running it locally rather than exposing a service is consistent with the rest of the product. The package manifest describes the application as AI-native video editing built around real creative timelines, generative workflows and local agent control, and the version in it matches the newest release tag, which is a small sign that releases and the manifest are kept in step. What is more telling is how the test scripts are organised, per feature rather than as one suite. There are separate scripts for media preparation, export scheduling, audio ducking, export readiness, export custom presets, interaction consistency, export workspace, multi-clip inspector and multi-clip effects, each running the native test runner over an explicit list of files. Two of them have a matching check script, and the export worker scheduling check even takes a serve flag, which means it starts something. A codebase with that many per-feature entry points has been debugged feature by feature rather than designed that way, which tells you more about the release history than a changelog would. The recent tags support that reading: smarter background media preparation, then an export scheduling hotfix, published the same afternoon, after an earlier release two weeks before.

Six installers, and the source archives are not one of them

Distribution is a GitHub Releases page with six assets and one instruction to ignore something you would otherwise assume was relevant. The assets are a Windows installer, a Windows portable build, a Mac build for Apple silicon, a Mac build for Intel, a Linux AppImage and a Linux Debian package. Two architectures for Mac is the detail worth noting, since it means a machine from the older generation is still supported rather than being quietly dropped. Then the instruction: ignore GitHub's auto-generated source-code archives unless you plan to build from source. That is a small courtesy that removes a real class of confusion, because the auto-generated archive is not a build of this application and installing it will not produce a working editor. The project is GPL-3.0 licensed, which for a desktop application distributed as binaries means the source obligation runs with the download, and it also means anyone shipping a modified build of it has obligations that a permissive licence would not impose. As of the latest tag the repository sits at a few hundred stars, a few dozen forks and a couple of dozen open issues, and it is not archived, which for a project at version 0.3.x with two releases on the same afternoon is a normal amount of churn rather than a sign of neglect.

Six translations and a backup summary left in the root

Two things in the repository root describe the project rather than the software. The documentation is translated into six languages besides English, with dedicated readme files for Spanish, Simplified Chinese, Japanese, Korean, Brazilian Portuguese and French, plus a page inviting help with the interface translation. That is a real localisation commitment for an application whose whole surface is buttons, panels and prompts. And there are two files describing the project: one summary document and a second one whose name begins with a backup marker and a number. A leftover backup file in the root of a repository is a small thing, but in a project whose package manifest is organised around per-feature discipline it is the kind of thing worth noticing, because it is exactly the sort of artefact that survives for years because nobody is looking for it. The rest of the tree is conventional for an Electron application with a build directory, an Electron directory, source, tests, infrastructure definitions, design mockups and a Linux build Dockerfile, plus agent configuration directories and a project MCP configuration file at the root, alongside the usual conduct, contributing, security, roadmap and release checklist documents.

Editorial conclusion

Use it if you already run ComfyUI locally and want the missing half of the workflow, because that is precisely the pitch: a production layer that plans the work, sends jobs to ComfyUI, collects the outputs and finishes the edit, instead of a second generation system. Do not adopt it expecting a single installer with everything in it, since all generation routes through a ComfyUI you run yourself on loopback. Three things to check first. Which ComfyUI you will point it at, since only local endpoints are supported and the default is a fixed port. Whether the guided creator workflows cover your case, since one of the four is described as experimental with rough edges. And which install artifact you take, since six are published and the source archives GitHub generates are explicitly not one of them.

Frequently asked questions

Official sources

1. License: GPL-3.0
2. Project website
3. README
4. Releases
5. VelornLabs/velorn on GitHub