OneTake

• 

  What OneTake does

• 

  Never run out of content

• 

  Before vs After OneTake

• 

  How your Agent Works

• 

  Hero section

Inspiration

Sarah sells houses. Her buyers find their agent on Instagram, so she knows she should post. But she spends her days in the car between showings, and the last time she tried to film a reel, a client called mid-sentence.

AI avatars were supposed to fix this. Instead they gave her a second job: five AI tools, a hundred prompts, copying between apps, and fixing a face that's almost hers. She doesn't want to become an AI operator. She wants her content done, and she wants it to sound like her.

There are millions of Sarahs: agents, stylists, coaches, consultants. For a one-person business, content is how new clients find you, and it's the first thing to go when the week gets busy.

What it does

OneTake is a content manager agent. Build once, create forever.

1. Build once (10 minutes): a few selfies and a few questions about your business.
2. Every week, talk for two minutes: record a voice note in the car, or type it. "Jess and Marco got their keys with 3.5% down. Jess cried in the driveway."
3. The agent does the rest. It finds the real stories in what you said, checks what your clients are reading this week, plans the week, and writes each post in your voice. Then it reviews its own work, and only after that does it make the posts with your face: a talking reel, a carousel and a photo post.
4. You get a page of finished posts, ready to post. No prompts, no tools, no editing.

It never puts words in your mouth. Every post is checked against what you actually said. If it invents a fact, it's rewritten. If it still fails, it's held back, and you aren't charged. In Sarah's live week on the judge account, the agent made her reel (9/10) and her photo post (10/10). It held back the carousel because the hook said "I thought I needed 20% down", and Sarah never said that.

Need something specific? One sentence in Create ("a photo of me handing keys to new homeowners") gets a finished photo, clip, reel or carousel.

How we built it

Every call that thinks or writes runs on NVIDIA Nemotron through Nebius Token Factory, sized to the job by one routing function (src/ai/text.ts):

• The agent (src/week/): listen → understand → research → plan → write → review → make. One stage runs per step under a lease, so it survives a closed tab, and a cron moves it along. Posts are written and reviewed in parallel.
• The reflection loop: write → critique (a score plus any invented facts) → rewrite with the critic's notes → keep the better draft, or hold the post back. Credits are spent only on posts that pass.
• Tavily searches this week's news. Nemotron Super keeps only results that matter to your audience, and the plan uses the topic, never a stranger's numbers.
• Gemini is used only where a text model can't go: images and video with your face, and transcribing voice notes.
• Next.js on Vercel, Clerk, Neon Postgres, private Blob storage. TypeScript strict, about 1,000 tests.

Challenges

Making the agent trustworthy, not just busy. In early live runs, the critic let a made-up "$200,000 city grant" through. Then it overcorrected and blocked posts for simply rewording her words. We tuned it until it blocks only facts she never gave, and caught it holding back a hook that put words in her mouth. We also learned to keep the better draft (one rewrite scored worse than the original) and to filter search results hard: a brand designer was once offered a football logo story.

On the open-model side, we gave reasoning enough token budget to finish, repaired empty JSON fields, and replaced typographic characters that fonts can't draw.

Accomplishments

• A two-minute voice note becomes a week of finished, on-brand posts with your face, end to end, on open models.
• A review loop that holds back posts with invented facts and reports what it removed, in plain words.
• A complete product: sign-up, setup, weekly agent, Create, library and credits. Live now.

What we learned

Nemotron's three sizes map cleanly onto an agent's jobs: Ultra thinks, Super writes and checks, Nano is quick. Token Factory's open-model pricing is what lets the agent afford a second opinion on every post, and that's what makes "done for you" something a business owner can trust.

What's next

Learning from the posts she approves, scheduling, and a mic in Create.

Built With

• clerk
• google-gemini
• nebius-token-factory
• neon
• next.js
• nvidia-nemotron
• tavily
• typescript
• vercel