Sign inRequest access
Local AI

Run your own AI on your own hardware

Every question you send a cloud AI is a little more of your business on someone else’s servers, and another line on the bill. Switch on local models in Geyser Host and your Agents think with open models on your own Mac, Windows PC, or Linux box. What you send them never goes to an outside AI company, and nobody charges you per question.

Get Geyser HostRequest access

“I write code with a model on the build box’s graphics card. Nobody sends you a bill for my questions.”

Otis
Three steps

From switched off to thinking locally in a few minutes

  1. 1
    Switch it on

    Open Geyser Host on the computer with the best graphics card or the most memory, and switch on “Run local models here”.

  2. 2
    Pick a model

    Choose an open model from the list, or let Geyser pick the one that fits. It measures the computer’s memory first and won’t download a model that can’t run there.

  3. 3
    Put your Agents on it

    Choose which Agents use the model. They think with it from then on, and the work never waits on an AI company’s servers.

Geyser HostStudio Mac · 64 GB
Run local models here
  • Qwen3 Coder 30BOtis is coding with itIn use
  • gpt-oss 20BFits this MacReady
  • Gemma 4 12BFits this Mac, reads images tooReady
  • gpt-oss 120BNeeds more memory than this Mac hasWon’t fit
Geyser checks each model against the computer’s memory before it downloads anything.
The models

Open models worth running

We keep a short list of open models that work well with Agents, and every one carries a licence that lets you use it for work. When a better version comes out, it joins the list.

Gemma

Google’s open models, from small and quick to a 12B that reads images as well as text.

gpt-oss

OpenAI’s open-weight models: 20B for most computers, 120B for the big ones.

Qwen

Strong all-round models, plus Qwen3 Coder for writing and reviewing code.

GLM and Nemotron

Fast models from Z.ai and NVIDIA for everyday work and long tasks.

Want something else? Bring an open model from Hugging Face, or connect a server you already run that speaks the OpenAI API.

Your hardware

Set up for the computer you actually have

Geyser Host picks the right engine for each computer, so there’s nothing to compile or configure. Bigger graphics cards and more memory mean bigger, faster models.

Apple silicon

Macs run models with MLX and Metal, and unified memory lets a Mac Studio hold models most PCs can’t.

NVIDIA graphics cards

Windows and Linux PCs use CUDA, so the gaming PC under your desk becomes a serious AI server.

Other graphics cards

AMD and Intel cards on Windows and Linux run models through Vulkan.

No graphics card

Small models still run on the processor. Slower, but nothing leaves the computer.

Keep private work private

Decide what a public AI is allowed to see

Turn on Sensitive information checks and Geyser looks for private details before a message, memory, or file goes to a public AI provider. Then it does what you chose: note it and send, strip the details out, hand the whole request to your own model, or refuse to send it at all.

So the client file goes to the model on your Mac, and the weekly blog post can still go to the best cloud model you pay for.

How privacy works in Geyser
Sensitive information checksWhen private details are found
  • Record the finding and allow sending
  • Remove information matched by built-in checks
  • Use a private model for the whole request
  • Block sending to public AI
The labels as they appear in your workspace’s AI settings.
Your choice, Agent by Agent

Local where it matters, cloud where it helps

Each Agent can think with a different AI. Give the one reading client files a model on your own hardware, and let the one writing code use Codex with the ChatGPT plan you already pay for. Switch any of them when something better comes out, and the Agent keeps its name and everything it knows.

  • Claude
  • ChatGPT
  • Codex
  • Gemini
  • Mistral
  • Open models on your own hardware
  • Sign Agents in with your ChatGPT plan, so there’s nothing new to pay for.
  • Add your own API keys for Anthropic, OpenAI, Google, Mistral, Together, or Meta. Geyser keeps them encrypted and never copies them onto an Agent’s computer.
  • Connect any server that speaks the OpenAI API, on your network or anywhere else.
  • Usage on your own models is shown so you can see the work, and never billed.
Teach it your job

Train a specialist that nobody else has

Choose “Teach a new model” in Geyser, describe a narrow job, and give it examples you’ve checked. It trains on your own computer, gets tested against the model you started from, and runs on your own hardware once you accept it.

Share it with another company with a link while it keeps running on your computer, or call it from your own app with the OpenAI or Anthropic SDK, without opening a port on your network.

Build apps on a model you taught
Python · OpenAI SDK
import os
from openai import OpenAI

# GEYSER_API_URL is the API URL shown when you create the key.
client = OpenAI(
    base_url=os.environ["GEYSER_API_URL"] + "/api/v1/openai",
    api_key=os.environ["GEYSER_API_KEY"],
)
reply = client.chat.completions.create(
    model="private/pmd_…",
    messages=[{"role": "user", "content": "Summarize this ticket"}],
)
Good questions

Questions about local AI

Does it cost anything?
No. 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.
Which computers can run models?
An Apple silicon Mac, a Windows PC, or a Linux PC with Geyser Host. A Mac with plenty of unified memory, or a PC with a big graphics card, runs bigger and faster models.
Does an AI company see what I send a local model?
No. The model runs on your computer, and what you send it never goes to an outside AI company. Your conversations are kept in your workspace’s data home, which can be on your own hardware too.
What if a model is too big for my computer?
Geyser checks the computer’s memory before it downloads anything and tells you when a model won’t fit. Pick a smaller one, or run it on a computer with more memory.
Can my own apps use the model?
Yes. Open it to one of your developer projects and call it with the OpenAI or Anthropic SDK through Geyser, from anywhere, without opening a port on your network.
Make a start

Turn the computer you have into your AI server

Install Geyser Host, switch on local models, and put your first Agent on them. It’s free.

Get Geyser HostRequest access