Robota uses ScreenCaptureKit to capture system audio (other meeting participants). This requires the "Screen & System Audio Recording" permission on macOS. Robota only captures audio - it never records your screen or takes screenshots. On macOS Sequoia (15+), Apple may ask you to re-confirm this permission monthly.

Robota mixes both audio streams (mic + system) 50/50 before transcription, which eliminates most echo. If you're still seeing duplicates, make sure you're using headphones - without them, your speakers bleed into your mic and can cause duplicates even with the mixing approach.

Apple's SpeechAnalyzer models are managed by the system and downloaded on demand. The first time you use a language, macOS needs to download the model (typically a few hundred MB). This happens in the background. If the download is slow, check your internet connection and try again - subsequent uses will be instant since the model is cached.

1. In Robota settings, set Summarization Provider to "On-device model"
2. Pick a model from the list and click Download. Models are roughly 1-5 GB and download once from Hugging Face, then stay on your Mac.
3. When the download finishes, click Use to make it active. Summaries now run entirely on-device.

This runs a full AI model inside Robota using a bundled copy of llama.cpp - no separate app or server to install, and fully offline after the one-time download. It's the easiest way to get a larger context window than Apple Intelligence (4K tokens) for longer meetings.

Custom models: under "Custom model" you can paste any GGUF from Hugging Face - the repo (for example bartowski/Qwen2.5-3B-Instruct-GGUF) and the .gguf file name - to use a model that isn't in the built-in list.

Already have a model file? Under "Local file" you can point Robota at a .gguf already on your Mac (from LM Studio, llama.cpp, or a manual download). It's used in place - no download, and nothing leaves your device.

1. Download and install Ollama
2. Pull a model: run ollama pull llama3.2 in Terminal
3. Make sure Ollama is running (it shows in your menu bar)
4. In Robota settings, set Summarization Provider to "Ollama"

Apple Intelligence is used by default and requires zero setup. For a larger context window without installing anything, try the "On-device model" provider above. Ollama is another option for users who already run it or want the largest context windows (16K tokens) for longer meetings.

1. Build llama.cpp (or install it via brew install llama.cpp)
2. Download a GGUF model from Hugging Face
3. Start the server: llama-server -m <model.gguf> -c 16384 --port 8080
4. In Robota settings, set Summarization Provider to "llama.cpp" and confirm the endpoint URL (default http://localhost:8080)

Robota talks to llama.cpp through its OpenAI-compatible /v1/chat/completions endpoint, so any GGUF model the server can run works. The context window is whatever you started the server with via -c.

Robota captures audio passively at the macOS system mixer level via ScreenCaptureKit. It does not inject into or interact with the meeting app in any way. Other participants cannot detect that Robota is running. Please ensure you comply with local laws regarding recording consent.

In Robota settings, set your Obsidian vault path and subfolder (defaults to Meetings/Robota). After transcription, click the "Obsidian" button in the review toolbar to export. You can also enable auto-save to export automatically after every meeting. Notes include YAML frontmatter, formatted transcript, summary (if generated), and bookmarks.

Robota requires macOS 26 (Tahoe) or later. This is because it uses Apple's SpeechAnalyzer framework for on-device transcription, which is only available on macOS 26+. Both Apple Silicon and Intel Macs are supported.

Audio files are written to a temporary directory and automatically deleted after transcription. Transcripts and summaries exist only in memory during the app session. Settings are stored at ~/Library/Application Support/com.worldtiki.robota/settings.json. If you export to Obsidian, markdown files are saved to your vault. No data is ever sent to any server. See the privacy policy for full details.