Odysseus local setup with oMLX and Gemma 4

This tutorial installs Odysseus directly on a computer so its AI agent can work with system files. The speaker explains that Docker's isolation limited those tasks in his earlier setup and recommends a separate machine for direct access. He follows the Apple Silicon installation path, while pointing to instructions for Linux and Windows.

The walkthrough covers cloning the repository, starting the macOS service, creating an admin login, and making the interface accessible over a trusted LAN or VPN. It also notes an optional clickable app wrapper, though the speaker keeps using the terminal.

For the local LLM connection, he uses oMLX to download and load a 12-billion-parameter Gemma 4 model, then enters the server URL and API key in Odysseus. He favors MLX over GGUF on Macs and discusses LM Studio and Ollama as other model tools. That performance preference is his assessment, rather than a benchmark established here.

The demonstration runs on an M5 Max with 128GB RAM. The speaker reports roughly 12GB for the model, then observes about 74GB RAM use during an agent task after setting a 256,000-token context. He asks it to inspect Downloads and recommend files to remove without deleting anything. An initial ls command fails, but the agent returns a response; browser access still needs setup. The closing discussion notes that Odysseus can download models through its cookbook, although the speaker cannot tell whether those files use MLX or GGUF.