# Bring your own AI

> Continue with ChatGPT on the plan you already pay for, use your own API keys, or run models on your computer.

Page: https://taalcut.app/ai/

Taalcut runs on the AI you already have: your ChatGPT plan, your own API keys, or models on your computer.

## Continue with ChatGPT

Sign in with OpenAI and Taalcut uses the ChatGPT plan you already pay for. There is no second bill from Taalcut.

- Accounts: In Settings › Accounts, the ChatGPT row offers Continue with ChatGPT.
- Sign in: OpenAI’s own page opens in your browser. You sign in there, and Taalcut never sees your password.
- Your plan: Back in Taalcut, the row says Using ChatGPT plan, with Manage usage and Sign out.

- **Where you sign in:** On OpenAI’s own page, in your browser. Taalcut never sees your password.
- **What it uses:** Your ChatGPT plan and its limits. Manage usage opens OpenAI’s page for them.
- **Where the sign-in lives:** Kept encrypted with a key in your computer’s keychain. It never reaches Taalcut’s windows, its library or its logs.
- **Signing out:** One button in Settings › Accounts.

## The AI it works with

Bring a key, point it at your own server, or use the models Taalcut runs on your computer.

### With your API key

- OpenAI: GPT models with your API key
- Anthropic: Claude models with your API key
- Google Gemini: Gemini models with your API key
- OpenRouter: Many models through one key
- ElevenLabs: Voice, speech and music

Paste a key in Settings › Accounts and Taalcut calls that provider for you. Taalcut keeps it encrypted with a key in your computer’s keychain, and never shows it back.

When you pay per use with a key, each new project starts with a budget of US$2. Taalcut checks it before paid work, and you can change it or turn it off.

### Your own server

- Ollama: Your own model server, in Settings › Accounts › Servers
- LM Studio: Your own model server, in Settings › Accounts › Servers

If you already run Ollama or LM Studio, add it in Settings › Accounts › Servers. You never need one: Taalcut already runs models on your computer.

### On your computer

- Taalcut’s own models: Search, text in video and music beats always run here; speech and the agent can too

## On your computer

Taalcut downloads and runs its own models. There is nothing to install, and no server to set up.

- **Always here:** Search, text in video and music beats. Your footage is indexed on your computer.
- **Speech:** Written down on your computer. It goes to ElevenLabs only if you add that key and allow it.
- **The agent:** Can run here: Qwen3.5 4B through llama.cpp on computers with 16 GB of memory or more.
- **On Ubuntu:** The agent on your computer needs a graphics card with 8 GB of video memory.
- **High-end Macs:** Qwen3.5 9B on MLX, on Apple silicon Macs with 32 GB of memory and a GPU with 16 or more cores.
- **8 GB Macs:** The agent runs a small text model. Understanding images needs a cloud model.

## Profiles and Local only

A profile sets the model for every job at once. Local only keeps everything of yours on your computer, whichever profile you picked.

### Profiles

A profile sets the model for every job at once. Pick one in Settings › Models.

- **Fast and cheap:** Quick, low-cost models. Good for first passes and simple edits.
- **Everyday:** Good results at a fair price. Right for most videos.
- **Best quality:** The strongest models. Slower, and each video costs more.
- **On this Mac:** Every job uses the models on this Mac. Lower quality, and no AI voice or images.

On Ubuntu the last profile is called On this computer, and its line says “this computer”.

### Local only

Turn on Local only in Settings › Privacy and nothing of yours leaves your computer, whichever profile you picked.

When it is on: Nothing of yours leaves your computer: no cloud models, uploads, reel links, feedback or usage counts. Update checks still send the app’s version and a random ID. Model downloads send nothing of yours.

### What leaves your computer while it is off

- **Cloud models you chose:** The agent’s messages and tool results, and video frames when a cloud model looks at a shot. Taalcut asks before it first sends your files to a cloud model.
- **Reference links:** Taalcut downloads a reel you paste, to read how it’s cut.
- **Speech:** Audio, only if you add an ElevenLabs key and allow it.
- **Crash reports and usage counts:** Only when you turn them on. Both start off.

Taalcut is in a private build. [Join the waitlist](https://taalcut.app/waitlist/) for one email when the beta opens.