Do I need to install .NET or Python?
No. The release archive is self-contained: the runtime, the speech engine, the voices and every component are already inside (~700 MB compressed). The only system packages the installer needs (curl, tar, libicu) are installed automatically when missing.
Which boards are supported? The Libre Computer Alta (5 TOPS NPU) and Solitude (1 TOP NPU, budget AI edge), running a 64-bit ARM Linux (Debian/Ubuntu-based or Armbian). The installer refuses to run on 32-bit ARM or x86.
Can I install it without an internet connection on the board? The one-line installer downloads the release from GitHub, so the board needs internet for the first install. After that, with a local model (Ollama / ExLlamaV2), the assistant works fully offline.
Where is it installed?
/opt/agentbridge (change with LIBRE_COMPUTER_AGENT_HOME). The service is called agentbridge and starts at boot.
How do I update? The assistant checks for updates by itself (see the Updates guide). You can also re-run the installer — your documents, keys and settings are never touched.
How do I chat with it?
The assistant is always on. For the full-screen terminal chat run sudo systemctl stop agentbridge then /opt/agentbridge/agent; restart the service when done. The HTTP API on port 5290 is available at any time to scripts and bots.
How do I choose which AI powers it?
Type /modelsetup in the chat: pick a local model (everything stays on the board) or a cloud provider (DeepSeek, Gemini, Anthropic) with anonymisation enabled by default.
Where are my documents? The assistant reads the documents folder it indexes on first start and answers questions from your files. You can ask it for finished DOCX, XLSX, PPTX and PDF documents.
Does text-to-speech work on the Libre Computer?
Yes — verified on ARM. The Kokoro neural voice, all voices and the ONNX model ship in the archive, and /v1/audio/speech produces WAV audio out of the box.
Does speech recognition (dictation / voice calls) work? Not yet on Linux ARM. The speech-to-text component is currently built for Windows only; the voice endpoints report themselves unavailable until an ARM build is published. Everything else — including the assistant speaking its answers — works.
Can the assistant answer phone calls? On Windows, yes (SIP). On the Libre Computer, the SIP bridge requires the speech-to-text component, so phone access is planned for a future ARM release. Telegram chat, in contrast, works fully.
Do my documents leave the device? With a local model, nothing leaves the board at all. With a cloud provider, the built-in anonymisation strips names and identifiers from the requests before they are sent.
Who can access the assistant? By default the HTTP API listens on localhost. Telegram has an allow-list; SIP has a PIN. Expose it to your LAN only if you need to, and read the privacy guide for the details.
The service won't start.
journalctl -u agentbridge -n 30 shows the reason. The most common causes are a missing libicu (the installer installs it when it can) or a malformed providers.json/appsettings.json.
Chat answers fail with "API key is not set".
No provider is configured. Run /modelsetup, or add the provider in providers.json next to the executable.
The installer says the architecture is unsupported.
The archive is built for 64-bit ARM. Check with uname -m — it must print aarch64.
Do I need a monitor? No. SSH into the board and run the installer; the assistant is reachable through the API on port 5290 from anywhere on your network.