Hi HN, I’m Jaime. I’m a VFX artist with over 20 years of experience, and I’ve been building Velorn as a solo developer.

Velorn is a GPLv3 desktop video editor for Windows, macOS, and Linux. It has a multi track video/audio timeline and works as a normal editor without any AI generation setup. You can import existing media, edit and layer clips, add effects, text and keyframes, transcribe locally with Whisper, mix audio, and export through FFmpeg.

For me, AI generated footage is raw material. The creative part is still what happens afterward: selection, timing, sound, pacing, effects, captions, and narrative structure. I wanted an app where generation and editing could exist in the same workflow, even editing live action footage with AI generations, so I started building Velorn.

One of the parts I've been especially interested in is MCP.

Velorn runs a local, loopback only MCP server that lets clients such as Codex and Claude Code interact with the project that's actually open in the editor.

Rather than putting a chatbot next to the timeline, I wanted to expose structured editing operations to the agent. It can inspect the project and timeline, look at media and frames, propose and perform edits, work with captions and effects, organize assets, and prepare exports and generation jobs.

Here's a short example using Codex:

https://www.youtube.com/watch?v=Owel8zkMWkY

For this demo I gave Codex almost no creative direction. I basically told it, you're live on YouTube, use Velorn to create a really cool motion graphics video. Just make something.

I wanted to see what would happen when the agent had access to an actual editing environment rather than asking it to execute a predetermined sequence of edits.

The video is about 3.5 minutes. I fast forward through some of the time where the agent is thinking, but otherwise it shows the interaction and the resulting work inside Velorn.

ComfyUI is not required for editing, captions, MCP, project management, or exporting. All generation in Velorn currently goes through a locally running ComfyUI instance on the same machine. For local models, your own GPU performs the work and Velorn doesn't charge credits. Generated results are brought back into the active project so they can immediately become part of the edit.

Some ComfyUI workflows can also use paid Partner Nodes/API services. Those may perform the actual computation in the cloud and require Comfy.org credentials and credits, but the workflow is still submitted through the user's local ComfyUI. Velorn doesn't currently operate a separate cloud generation backend. So someone can download Velorn and use it as an editor with MCP without configuring ComfyUI at all. ComfyUI is only required when using Velorn's current generation features.

Velorn uses Electron, React, Zustand for project/timeline state, FFmpeg for media processing and export, local Whisper for transcription, ComfyUI's API for generation, and a local MCP server exposing structured editorial operations.

I chose Electron because it gave me one cross platform editor implementation while still allowing integration with the filesystem, FFmpeg, local processes, ComfyUI, and MCP.

The project has picked up 400+ GitHub stars so far as a solo developer, I'm particularly interested in hearing from people who actually edit video or are experimenting with agents.

I would really love some feedback.

Velorn is free and open source. Try it!

To see longer MCP interactions without the agent's thinking being fast forwarded, I also recorded two uncut examples:

https://www.youtube.com/watch?v=_r4jf7ZDT2o

https://www.youtube.com/watch?v=AT9usQS3m48

Here's a music video that I made in just 2 days with Velorn that has gotten some attention:

https://www.youtube.com/watch?v=iX-YdjVMDhg

GitHub: https://github.com/VelornLabs/velorn Website: https://velorn.ai

Thanks for looking!

Show HN: Bento Term – native tmux client with agent states and voice input

A controversial Alzheimer's surgery is said to reverse symptoms

Briefing on Extinction-Level AI Threats

Relentlessly Resourceful (2009)

The Book Is a Book, Not Disembodied Information

DeepSeek Harness Desktop Version

Show HN: Clair APIs – an active jobs dataset from public ATS boards

The paradox of the perfect code editor (2021)

US to tell partners they must pick sides in AI race with China

Godfather of AI: "If you sleep well tonight..." [video]

Is AI Coding Effecting Our Mental Health? [video]

Socync – Automatic social media posting and scheduling

Qwen 3.8 27B at 2x speed on a 5090

We Need to Talk About Leopold [video]

I built Fusekit, a place to deploy internal web apps

A 25-year-old video patent just expired, ending a legal headache for Linux

How to get a matmul kernel into the top percentile (worklog)

Show HN: Bluetooth keyboard and Windows 95 cursor on reMarkable

The origin of vertebrate teeth and evolution of sensory exoskeletons

Find duplicate code across nine languages

Show HN: Velorn – an open-source desktop video editor with MCP agent control

Show HN: MyFirst.News – AI-generated podcast for curious kids

Ask HN: You have model orders of magnitudes faster than transformer, so?

Humanists Have Done Nothing Wrong

JOE – Joe's Own Editor

Show HN: Ansi.md – Everything I know about building modern CLI apps

What makes a good puzzle image?

One agent, every surface: how we built the Kiro agent harness

Apple takes down billboard after accusations of 'electronic babysitting'

The price of a Costco hot dog has gone up