Where it runs
Geyser Host picks the right engine for the computer: CUDA on NVIDIA cards, Vulkan on AMD and Intel cards, and MLX and Metal on Apple silicon Macs.
A gaming PC with a big NVIDIA card, or a Mac Studio with plenty of memory, spends most of its day doing nothing. Install Geyser Host, switch on “Run local models here”, and pick a model that fits. Agents on your other computers can think with it too, for as long as the PC is on, and nobody bills you per question.
“I live on your MacBook and think with the model on the gaming PC, so nobody bills you for my questions.”
Your Agents answered 1,900 questions last month with the graphics card in the gaming PC, and none of them showed up on an AI bill.
Installs Geyser Host on the PC with the RTX card and switches on “Run local models here”.
Chooses gpt-oss 20B. Geyser checks the PC’s memory first, so it downloads only a model that fits.
Moves from a cloud model to gpt-oss on the gaming PC, and keeps its name and everything it knows.
Reads the week’s newsletters and leaves you a one-page brief, with the thinking done on the PC’s graphics card.
“Otis still writes code with a cloud model. Qwen3 Coder 30B would fit on the gaming PC too. Want me to show you how to switch Otis over?”
Geyser Host picks the right engine for the computer: CUDA on NVIDIA cards, Vulkan on AMD and Intel cards, and MLX and Metal on Apple silicon Macs.
Agents on any of your computers can use the model while the PC is on. Put the ones that read private files on it, and keep a cloud model for the jobs that need one.
You choose which Agents use the model, and Geyser won’t download a model that can’t run on the computer.
Models on your own computers work on every plan, including Free, and Geyser doesn’t bill for them. Beyond the hardware, your only cost is electricity.
Install Geyser Host on the computer with the best graphics card or the most memory.
Switch on “Run local models here” and pick a model, or let Geyser pick one.
Choose which Agents use it.
Claude Code on your laptop and Codex on the build box keep going while you’re out. Your Agents tell your phone which session finished and which one is waiting on you.
Read the storyOtis codes with Qwen3 Coder on the gaming PC’s graphics card. The repo stays on your machines, and the model has no per-token bill.
Read the storyAda reads your tax papers and medical bills with a model on your own Mac, so none of it goes to an outside AI company.
Read the storySet up a workspace, give an Agent something useful to do, and see how far it gets.