Launch HN: Mosaic (YC W25) – Agentic Video Editing

[adishj]

Hey HN! We’re Adish & Kyle from Mosaic (https://edit.mosaic.so, https://docs.mosaic.so/, https://mosaic.so). Mosaic lets you create and run your own multimodal video editing agents in a node-based canvas. It’s different from traditional video editing tools in two ways: (1) the user interface and (2) the visual intelligence built into our agent.

We were engineers at Tesla and one day had a fun idea to make a YouTube video of Cybertrucks in Palo Alto. We recorded hours of cars driving by, but got stuck on how to scrub through all this raw footage to edit it down to just the Cybertrucks.

We got frustrated trying to accomplish simple tasks in video editors like DaVinci Resolve and Adobe Premiere Pro. Features are hidden behind menus, buttons, and icons, and we often found ourselves Googling or asking ChatGPT how to do certain edits.

We thought that surely now, with multimodal AI, we could accelerate this process. Better yet, an AI video editor could automatically apply edits based off what it sees and hears in your video. The idea quickly snowballed and we began our side quest to build “Cursor for Video Editing”.

We put together a prototype and to our amazement, it was able to analyze and add text overlays based on what it saw or heard in the video. We could now automate our Cybertruck counting with a single chat prompt. That prototype is shown here: https://www.youtube.com/watch?v=GXr7q7Dl9X0.

After that, we spent a chunk of time building our own timeline-based video editor and making our multimodal copilot powerful and stateful. In natural language, we could now ask chat to help with AI asset generation, enhancements, searching through assets, and automatically applying edits like dynamic text overlays. That version is shown here: https://youtu.be/X4ki-QEwN40.

After talking to users though, we realized that the chat UX has limitations for video: (1) the longer the video, the more time it takes to process. Users have to wait too long between chat responses. (2) Users have set workflows that they use across video projects. Especially for people who have to produce a lot of content, the chat interface is a bottleneck rather than an accelerant.

That took us back to first principles to rethink what a “non-linear editor” really means. The result: a node-based canvas which enables you to create and run your own multimodal video editing agents. https://screen.studio/share/SP7DItVD.

Each tile in the canvas represents a video editing operation and is configurable, so you still have creative control. You can also branch and run edits in parallel, creating multiple variants from the same raw footage to A/B test different prompts, models, and workflows. In the canvas, you can see inline how your content evolves as the agent goes through each step.

The idea is that canvas will run your video editing on autopilot, and get you 80-90% of the way there. Then you can adjust and modify it in an inline timeline editor. We support exporting your timeline state out to traditional editing tools like DaVinci Resolve, Adobe Premiere Pro, and Final Cut Pro.

We’ve also used multimodal AI to build in visual understanding and intelligence. This gives our system a deep understanding of video concepts, emotions, actions, spoken word, light levels, shot types.

We’re doing a ton of additional processing in our pipeline, such as saliency analysis, audio analysis, and determining objects of significance—all to help guide the best edit. These are things that we as human editors internalize so deeply we may not think twice about it, but reverse-engineering the process to build it into the AI agent has been an interesting challenge.

Some of our analysis findings: Optimal Safe Rectangles: https://assets.frameapp.ai/mosaicresearchimage1.png Video Analysis: https://assets.frameapp.ai/mosaicresearchimage2.png Saliency Analysis: https://assets.frameapp.ai/mosaicresearchimage3.png Mean Movement Analysis: https://assets.frameapp.ai/mosaicresearchimage4.png

Use cases for editing include: - Removing bad takes or creating script-based cuts from videos / talking-heads - Repurposing longer-form videos into clips, shorts, and reels (e.g. podcasts, webinars, interviews) - Creating sizzle reels or montages from one or many input videos - Creating assembly edits and rough cuts from one or many input videos - Optimizing content for various social media platforms (reframing, captions, etc.) - Dubbing content with voice cloning and lip syncing.

We also support use cases for generating content such as motion graphic animations, cinematic captions, AI UGC content, adding contextual AI-generated B-Rolls to existing content, or modifying existing video footage (changing lighting, applying VFX).

Currently, our canvas can be used to build repeatable agentic workflows, but we’re working on a fully autonomous agent which will be able to do things like: style transfer using existing video content, define its own editing sequence / workflow without needing a canvas, do research and pull assets from web references, and so on.

You can try it today at https://edit.mosaic.so. You can sign up for free and get started playing with the interface by uploading videos, making workflows on the canvas, and editing them in the timeline editor. We do paywall node runs to help cover model costs. Our API docs are at https://docs.mosaic.so. We’d love to hear your feedback!

[primitivesuave]

Last year, I made a YouTube documentary series showcasing the prolific corruption in a small city government. I downloaded all the city government meetings, used Whisper to transcribe them, and then set up a basic RAG so I could query across a decade of committee meetings (around 1 TB of video). Once I got the timestamps that I'm interested in, I then have to embark on a tedious manual process of locating the file, cutting out a few seconds/minutes from a multi-hour video, and then order all the clips into a cohesive narrative.

These seem like problems that LLMs are especially well-suited for. I might have spent a fraction of the time if there was some system that could "index" my content library, and intelligently pull relevant clips into a cohesive storyline.

I also spent an ungodly amount of time on animations - it felt like "1 hour of work for 1 minute of animation". I would gladly pay for a tool which reduces the time investment required to be a citizen documentarian.

[adishj]

hey, thanks for sharing about your documentary series. would love to check it out if you don't mind linking it!

we don't yet support that volume of footage (1TB), however if you'd like to try this at a smaller scale, you can already do this today with the Rough Cut tile — simply prompt it for the moments that you're interested in (it can take visual cues, auditory cues, timestamp cues, script cues) and it will create an initial rough cut or assembly edit for you.

I'd also recommend checking out the new Motion Graphics tile we added for animations. You can also single-point generate motion graphics using the utility on the bottom right of the timeline. Let me know if you have any questions on that.

[breadislove]

you should check mixedbread out. we support indexing multimodal data and making data ready for ai. we are adding video and audio support by the end of the year. might be interesting for the OP as well.

we have couple investigative journalists and lawyers using us for a similar usecase.

[robotswantdata]

Gemini 3 would rip through that problem, but equally you could slice the video with existing open source tooling such as FFMPEG then combine with blender for the video curation. Gemini 3 could probably write you the workflow as well.

[cjbarber]

I think this is a great endeavor. I was thinking about a channel that I like watching on YouTube. They travel to exotic places by boat and film themselves, nature documentary style. To make good videos requires going to these places, a ton of filming, AND a ton of editing. They put out a video every 2 weeks or so on their trips. I imagine the editing is the hard part.

This is a long winded way of saying that I think creators need what you're making! People who have hours of awesome footage but have to spend dozens of hours cutting it down need this. Then also people who have awesome footage but aren't good at editing or hiring an editor, same thing. I'd love to see someone solve this so that 90th percentile editing is available to all, and then it can be more about who has the interesting content, rather than who has the interesting content and editing skills.

[adishj]

thanks! Mosaic can already do the rough cuts for you — so you can upload all your footage from your travel, and prompt it to "make a 2 minute highlight reel of your trip to Japan", for instance.

soon, we also plan to incorporate style transfer, so you could even give it a video from the channel you enjoy watching + your raw footage, and have the agent edit your footage in the same style of the reference video.

[moinism]

Hey, this is super cool. congrats on the product and the launch!

I'm building something exactly similar and couldn't believe my eyes when I saw the HN post. What i'm building (chatoctopus.com) is more like a chat-first agent for video editing, only at a prototype stage. But what you guys have achieved is insane. Wishing you lots of success.

to healthy competition!

[ack210]

I just signed up for a Creator plan, but it looks like the automated "Thank you for being a Mosaic Creator" email going out is not configured correctly. Instead of having my company name, it referenced a different business name and description (that seems to exist/be accurate, so not a placeholder).

[adishj]

Hey! Thanks for calling this out — looking into what happened here & fixing right now.

[adishj]

This has been fixed now.

[djeastm]

Move fast, break things... like privacy regulations.

[ansc]

Woah, yikes.

[Forgeties79]

> We got frustrated trying to accomplish simple tasks in video editors like DaVinci Resolve and Adobe Premiere Pro. Features are hidden behind menus, buttons, and icons, and we often found ourselves Googling or asking ChatGPT how to do certain edits.

Hidden behind a UI? Most of the major tools like blade, trim, etc. are right there on the toolbars.

> We recorded hours of cars driving by, but got stuck on how to scrub through all this raw footage to edit it down to just the Cybertrucks.

Scrubbing is the easiest part. Mouse over the clip, it starts scrubbing!

I’m being a bit tongue in cheek and I totally agree there is a learning curve to NLE’s but those complaints were also a bit striking to me.

[adishj]

hey! You're right that most of the basic tools like splitting / trimming are available right in the timeline. but things like adding a keyframe to animate a counter, for instance, I had no idea where to go or how to start.

Scrubbing is easy enough when you have short footage, but imagine scrubbing through the footage we had of 5 hours of cars driving by, or maybe a bunch of assets. This quickly becomes very tedious.

[andrewmlevy]

obligatory https://news.ycombinator.com/item?id=9224

[kul]

Can it work for this use-case? I have lots of 15 seconds to 1 min duration videos) of my kids and want to upload them all (let's say 10 videos) and have the agent make a single video with all the best bits of them?

[adishj]

yes! you can upload as many videos as you want (file limits currently are at 20GB and 90 minutes, per file). then I'd recommend using either the Rough Cut tile or the Montage tile to stitch them all together. In those tiles, you can prompt particular visual cues in terms of how you want the videos to be combined. Let me know if any questions.

[sails]

I’ve had a lot of fun with Remotion and Claude Code for CLI video editing. I’ve been impressed with how much traditional video editing I can manage.

I will be checking this out!

[adishj]

that's super interesting — what kind of things have you done with remotion and Claude Code?

they're very powerful, when you put them together, it almost feels like Cursor for Video Editing

[hypnagogicjerk]

Interested in your workflow @sails

[shraey_92]

I like the tile-based workflow approach. I’m curious, is integration with tools like 11labs/cartesia or HeyGen on the cards? It would make it much easier to produce influencer-style POV/first-person content using digital avatars and cloned voice-overs.

Also, do you have an API available to trigger workflows programmatically?

[homeonthemtn]

These comments real sus.

[dang]

Yes, a bunch of positive comments from accounts without much posting history appeared in this thread. I assume that the OP's friends got wind of their launch.

We tell founders to avoid that (scroll down to the bold part of https://news.ycombinator.com/yli.html for how we try to scare YC founders into not doing it!) - but to be fair, (1) this is not always easy to control, and (2) people posting such comments think they're helping and don't have enough experience of HN to realize that it has a counter-effect.

I'm going to move the overly sus ones to a collapsed stub now. (https://news.ycombinator.com/item?id=45988584)

[adishj]

i agree, things are a bit too kind. give me some more feedback.

[rsancheti]

And what’s the plan for determinism? For repeat workflows it’s important that the same pipeline produces the same cut each time. Are node outputs consistent or does the model vary run to run?

[adishj]

since we're building on top of LLMs which are by nature probabilistic, you won't produce the exact same frame-level cut each time, but of course there is still determinism in the expected outputs

for example, if you have a workflow setup to create 5 clips from a podcast and add b-rolls and captions and reframe to a few different aspect ratios, any time you invoke this workflow (regardless of which podcast episode you're providing as input), you'll get 5 clips back that have b-rolls, captions, and are reframed to a few different aspect ratios

however, which clips are selected, what b-rolls are generated, where they're placed — this is all non-deterministic

you can guide the agent via prompting the tiles individually, but that's still just an input into a non-deterministic machine

[jaccola]

Very cool. It definitely feels to me that the power of pro tools should be available to more people with AI.

Would have been nice if there was a killer demo on your landing page of a video made with Mosaic.

[adishj]

that's our perspective as well.

a lot of tooling is being built around generative AI in particular, but there's still a big gap for people that want to share their own stories / experiences / footage but aren't well-versed with pro tools.

valid feedback on the landing page — something we'll add in.

[bluelightning2k]

The problem is, any video demo of a tool like this is just an entirely unrelated video.

[zkmon]

I just clicked the link and encountered a non-scrollable, dark, fixed content pane with loads of flickering images and scrolling text with random font sizes without much meaning. I felt imprisoned, subjected to unexpected suffering, can't scroll away, got scared and raced for the window close button, and then breathed easy.

[adishj]

seems like the landing page is detracting from the main product, this is good feedback so thanks! For now, avoid the scaries and head directly to https://edit.mosaic.so to try the actual canvas interface

[pelagicAustral]

They really managed to handcraft a unique user experience, that's for sure.

[deepspace]

I had the same reaction. About what you would expect from a team steeped in the Tesla mindset.

[vivzkestrel]

did not get this part "After talking to users though, we realized that the chat UX has limitations for video: (1) the longer the video, the more time it takes to process. Users have to wait too long between chat responses. (2) Users have set workflows that they use across video projects. Especially for people who have to produce a lot of content, the chat interface is a bottleneck rather than an accelerant." what are you processing? frame by frame images?

[adishj]

what I mean here is that processing / analyzing a beefy file format like video will take longer than processing text input

same with returning that back to the user as manipulated output (text / code generation is much more rapid than rendering a video)

[ttoinou]

This website seems to be coded with AI. The sign up only works in Chrome not Chromium, and then I'm stuck forever at the step 3 of signing up

[dwrodri]

Have y'all talked with Max and the Ozone team? Suppose you would have lots to learn from them as you take on this space. Best of luck, video is hard!

[adishj]

Haven't chatted with them but their platform looks interesting!

Video is hard, but it's also a fun modality which presents some interesting challenges. And is where content is converging towards.

[sbfeibish]

The home page is a big turn off. Change it.

[PeakRhinoceros]

I found it difficult to understand what Mosaic is based off the home page. It would be helpful to synthesize the overview from the docs so visitors don’t have to navigate around to find out what the tool is. Providing some screenshots would be good too, ideally with a video showcasing how the tool works.

[djeastm]

That's not very constructive criticism. Change it in what way(s)?

[HanClinto]

I absolutely love your approach of "expert tools". If I understand your approach, you aren't just feeding a video into a multimodal LLM and asking it "what is the bounding box of the optimal caption region?" -- you have built tools with discrete algorithms (using traditional CV techniques) that use things like object detection boxes + traditional motion analysis techniques to give "expert opinions" to the LLM in the form of tool calls -- such as finding the regions of minimal saliency + minimal movement to be the best places for caption placement.

If the LLM needs to place captions, it calls one of these expert discrete-algorithm tools to determine the best place to put the captions -- you aren't just asking the LLM to do it on its own.

If I'm correct about that, then I absolutely applaud you -- it feels like THIS is a fantastic model for how agentic tools should be built, and this is absolutely the opposite of AI slop.

Kudos!

[adishj]

thanks for the comment, thats exactly right

we're using a mix of out-of-the-box multimodal AI capability + traditional audio / video analysis techniques as part of our video understanding pipeline, all of which become context for the agent to use during its editing process

[callamdelaney]

Hey, good luck with Mosaic.

Some feedback initially on the landing page, looks great but I thought that there is, for me, too much motion going on on the homepage and the use cases page. May be an unpopular opinion!

[cjbarber]

Agreed, homepage was confusing for me also. I tried to scroll around and see a demo. For a product like this that is so visual, I expected to be able to find a 30s demo clip somewhere but couldn't see one on the homepage or product page (and the scrolling on the product page was annoying for me).

[echelon]

Can you make this a desktop app?

I'm really tired of editing videos in the cloud. I'm also also tired of all these AI image and video tools that make you work over a browser. Your workflow seems so second class buried amongst all the other browser tabs.

I understand that this is how to deploy quickly to customers, but it feels so gross working on "heavy" media in a browser.

[supportengineer]

There's plenty of great native desktop apps for video editing. And there have been for almost 30 years. I also don't understand why anyone would want to use a browser for this.

[adishj]

we've done a ton of work to optimize the uploads / downloads / transcoding of videos to handle beefy files using proxies, and also allow you to XML export back to traditional editing tools that can link back to your "heavy" media, but I hear you and I think anything running locally on device is just going to feel faster

it does present its own set of challenges, but something we've thought about

[Tetraslam]

this is going to save me so much time, hell yeah guys!

[adishj]

thank you! let us know if you have any feedback!