Skip to content

Ollama Backend

Local vision models via Ollama. Fully private, no API key needed.

Setup

  1. Install Ollama: ollama.com
  2. Start the server:
ollama serve
  1. Configure eyeroll:
export EYEROLL_BACKEND=ollama

Or run eyeroll init and select Ollama.

Auto-install

On macOS and Linux, if Ollama is not installed, eyeroll will attempt to install it automatically and start the server. This happens on first use.

Capabilities

Feature Supported
Direct video upload No
Frame-by-frame analysis Yes
Audio transcription No
Text generation Yes

Models

Default model: qwen3-vl

Available vision models for Ollama:

Model Size RAM needed Quality
qwen3-vl ~5GB 8GB+ Good
qwen3-vl:2b ~1.5GB 4GB+ Acceptable
qwen3-vl:8b ~5GB 8GB+ Good

Override with the --model flag:

eyeroll watch video.mp4 --backend ollama --model qwen3-vl:2b

Auto-pull

If the specified model is not installed, eyeroll pulls it automatically on first use. Progress is shown in stderr.

How it works

Ollama only supports image analysis, not direct video or audio. eyeroll always uses the frame-by-frame strategy:

  1. Extract key frames from the video
  2. Send each frame as a base64-encoded image to the Ollama API
  3. Audio transcription is skipped entirely
  4. Synthesize the report from frame analyses only

Configuration

Environment variable Default Description
EYEROLL_BACKEND gemini Set to ollama to use this backend
OLLAMA_HOST http://localhost:11434 Ollama server address

Limitations

  • No audio transcription -- narration in videos is lost
  • Frame-by-frame only -- no motion or timing context
  • Quality depends on model size and hardware
  • Slower than cloud backends on most hardware
  • qwen3-vl:8b has known issues reading screen recordings -- use the default qwen3-vl or qwen3-vl:2b instead
  • qwen3-vl:2b can be slow on the synthesis step for longer videos

Silent recordings

Most developer screen recordings are silent, so the lack of audio transcription is often not a problem. But if the video has spoken narration, important context will be missed with Ollama.

Cost

Free. All processing happens locally.

Install

pip install eyeroll

No extra SDK needed -- Ollama communication uses the standard library (urllib).