NewLocal AI on a graphics card in your company

One brain for the whole company. Everyone sees what they should.

A company second brain with role-based access. Everyone works with it in their own Claude, ChatGPT or Codex, or in the company's own AI, and gets only the folders that are theirs.

We'll only use your e-mail to get in touch.

You pay per person, not for another AI licence.

Claude CodeCodexObsidian

Click a role to see what it gets from the brain.

The problem

A Git repo is all or nothing

A second brain in a folder works great while only developers use it. As soon as you want to share it with more people, you hit a wall.

  • You can't share just a part

    Sales would see salaries. A contractor, the whole company.

  • Git is a wall for people

    The accountant won't push or set up a terminal.

  • You pay for AI twice

    Other tools bundle their own model into the licence, although you already pay for Claude.

Works with Claude, ChatGPT and Codex

Company knowledge in one place. Everyone gets their part.

The brain is a folder with what the company knows. Roles decide who sees what, and everyone works with it in the AI they already use.

An example

The same question, two roles

The brain knows where to look for the answer. And it knows what a given person must not see.

What did we charge Acme for maintenance last year, and what does the contract say about response times? Maintenance at CZK 12,000 a month, invoiced quarterly. Response within two working days, within 4 hours for an outage. (contract-acme-2025.pdf, p. 3)

Meanwhile sales does not see salaries or leadership's notes. Their role does not have them, so neither does their AI.

Why a brain

AI makes you 10× faster. A brain makes it far more.

AI on its own is smart, but it knows nothing about your company. The brain gives it memory, context and rules.

  1. 1×: Without AI

    You search e-mails, documents and colleagues' heads. Know-how leaves with people.

  2. 10×: With Claude or ChatGPT

    The AI writes and advises, but knows nothing about the company. You explain everything again every time.

  3. far more: AI + onlybrain

    The AI knows clients, prices and processes. Agents do the repetitive work and everyone sees only their part.

AI your way

The brain works with the AI you choose

No need to switch to another AI. People keep their own tools, and the company can connect its own model or its own key. The role's permissions apply the same way every time.

  • The subscriptions you have

    The AI people already use

    Claude, ChatGPT and Codex connect to the brain with a connector (MCP), and Claude Code and Codex can also open it as a folder. Everyone sees only what their role allows.

  • Your server

    Your own AI on your server

    Your own Claude through a runner: it runs at your place and only connects out from the server, you open nothing inward. Or a model with an OpenAI-compatible API, such as GPT, vLLM or Ollama.

  • Your key

    Your own key

    The chat and other AI features in the app run on your company's Anthropic key, at the same quality. You pay the provider directly and know where the data goes.

An admin decides which feature runs on which AI and sees the jobs and the usage. The AI only ever gets what the role may see. For now we connect a company's own AI together with the first companies in a pilot.

What about Perplexity, Gemini or Copilot?

The brain is a plain folder and an MCP connector. We only list what really works today.

works

Connector (MCP)

On the web, on the computer and on the phone, also in code editors and the terminal. Sign in with onlybrain, permissions by role.

  • Claude
  • ChatGPT
  • Claude Code
  • Codex
  • Cursor
  • VS Code with GitHub Copilot
  • Gemini CLI
works

Folder on your computer

The onlybrain app keeps the brain as a plain folder that any tool can open.

  • Cursor
  • Windsurf
  • VS Code with GitHub Copilot
  • Obsidian
coming

Other big tools

They already support custom MCP connectors, we are now checking the onlybrain sign-in with them.

  • Perplexity
  • Gemini Enterprise
  • Microsoft 365 Copilot
  • Mistral Le Chat
Local AIIn pilot

AI on a graphics card in your company

The model runs on your own card and you control it right inside onlybrain. The company sets which jobs run only on its card. When the card can't keep up, a job waits, fails or is taken over by the fallback in the onlybrain cloud, depending on the settings. With the fallback off, data never goes to a third-party AI provider, and with Enterprise the whole operation stays in the company.

Arrange a pilot

We are fine-tuning local AI with the first companies in a pilot. We measure the speed on your hardware together and won't promise numbers in advance.

An illustration, not a screenshot. No measured numbers.

Five modes, one card

  • One person

    The largest model and a long context for the best answers for a single user.

  • Team

    Settings for several people at once: colleagues share the card, not just one person.

  • Agent

    A model tuned for reliable tool calls, for the jobs onlybrain does for you.

  • Transcription coming

    Meeting transcription and a small model beside it, both fit in the card memory. Transcription on your own card is not in operation yet.

  • Images and video coming

    The whole card memory for generating images and video.

The card does one mode at a time. You switch it in onlybrain with one click.

  • The model you want

    The default is an open Qwen model. Pick others in the catalog with labels (Czech, tools, licence) and download them with one click. For customers only Apache 2.0 or MIT licences.

  • Memory calculator

    From the model metadata it works out what fits on the card: with what context and for how many people. The sliders show right away what changes.

  • Measure and tune

    One button tries the settings on your card and keeps the best. You see the speed measured on your own hardware, not from a brochure.

  • Mode calendar and job queue coming

    Team during work hours, something else otherwise. Jobs the card cannot take right now wait in the queue. Not deployed yet.

  • One installation

    One package (the runner) goes on the server with the card. It connects outward by itself, you open nothing into the company.

  • Controlled in onlybrain

    Modes, models, card status and logs are in the organization settings. No other site or tool to add.

How it works

  1. Connect the server

    Install the runner on the server with the graphics card and pair it with onlybrain by a one-time code. It works with the onlybrain cloud and with Enterprise.

  2. Choose a mode

    In onlybrain you pick what the card does and the model. The calculator tells you what fits.

  3. Work in onlybrain

    Summaries and file descriptions run on your card, meeting transcripts on the card are coming. You turn the cloud fallback on or off according to what the company may do.

AI rolesComing soon

Coming soon: AI roles for sensitive data

Roles say who sees what. AI roles are meant to say which AI may get which data: you will be able to mark a folder "local AI only" and onlybrain will send nothing from it to a cloud AI.

Not deployed yet. This is a plan, not a feature to rely on today. We will announce it when it is in production.

An illustration of the planned feature, not a screenshot.
  • Any AI

    The default. With no setup nothing is meant to change, everything runs as today.

  • Company AI only

    Content will be allowed only to local AI, the company's own key or its own server. Not to the onlybrain account.

  • Local AI only

    Content will be allowed only to the graphics card inside the company. Cloud AI is not meant to get it at all.

  • The server is to enforce it, not a prompt

    Content an AI must not get is not meant to be visible to it at all. We don't rely on the model obeying.

  • A subfolder can only tighten it

    The strictest label from the folder and all its parents is to apply. No exceptions downward.

  • A minimum for the whole company

    The organization admin sets a floor that no brain can go below.

  • An audit without content

    It will record which function, which AI and how many paths, not what was in them.

Once done, it will guard the AI inside onlybrain. It won't stop a person who sees the folder through their role from opening a copy in their own tool.

Who it is for

Local AI and AI roles are aimed at companies that want AI but do not want to hand part of their data to a third-party provider.

  • Accounting and tax firms

    You want client data on your own card.

  • Companies under NDA and law firms

    You don't want to send contracts and correspondence to a third-party AI vendor.

  • Manufacturing with know-how

    Drawings, recipes and procedures must not reach a competitor or a third-party cloud.

In numbers

Clear rules, no surprises

3ways to bring AI: the subscriptions people have, the company's server, or its key
3access levels for every folder: hidden, view, edit
0hidden files in a role's copy, not even in the history
4kinds of brain: company, department, work at a company, personal
Security

Your knowledge stays yours

The gateway checks access before anything is downloaded. A hidden folder is not in the clone, the history or the commit messages.

  • Encrypted connection

    Clones, changes and the app only travel over encrypted HTTPS.

  • We don't look inside

    The onlybrain team does not look into brain contents, only at your explicit request.

  • Hidden means missing

    What a role can't see is not in its copy at all. Not even its AI can reach it.

  • Audit

    Every access and every change is in the audit log.

Enterprise

All of onlybrain inside your company

The app, the documents, the history of changes and the AI itself run on your servers. With a local model, neither the data nor the AI queries leave the company.

Arrange a pilot or a consultation

We are fine-tuning Enterprise with the first companies in a pilot. We price it to fit you.

  • Local modelvLLM · Ollama
  • Sign-inEntra ID · Keycloak
  • People and rolespermissions per folder
  • Documentsfiles and history
  • On your server

    onlybrain and the AI run at your place, on a server or a virtual machine in your network.

  • Local models

    The model runs at your place, for example with vLLM or Ollama. Or connect the AI you already run.

  • Company sign-in

    People sign in with their company account through Entra ID, Keycloak or another OIDC provider.

  • Your look

    Your company's name, logo and colours, without touching the code.

  • Licence without internet

    The licence is checked right on your server, with no connection to us.

  • Roles and audit as in the cloud

    Permissions per folder, approving changes and the access log work the same way.

Two options

  • On your server

    Installed on your server or virtual machine, with a local model, or connected to the AI you already run.

  • Private instance with us in the Czech Republic

    A separate instance just for your company on a server in the Czech Republic. We take care of running and updating it.

Pricing

Every plan adds one clear thing

Have as many brains as you like. You pay per person.

  • Free

    A taste
    • 1 small brain
    • basic template
  • Pro

    A full brain for yourself
    • capabilities, agents, skills
    • desktop app
    • marketplace
  • Team

    Sharing whole brains
    • everything in Pro
    • team member management
  • Business

    for companies
    Sharing parts of a brain
    • roles per folder
    • audit and SSO
    • sub-brains as a live view

Enterprise

Tailored

All of onlybrain inside your company, or a private instance with us in the Czech Republic.

Arrange a consultation

Prices will be published at launch. We also offer a tailored setup: we learn how your company works and build the brain for you.

Questions

Frequently asked questions

Do colleagues need to know Git?

No. Who works in Claude Code or Codex gets the brain with one command and the AI handles Git for them. Everyone else gets an app in which the brain is an ordinary folder that syncs by itself.

Who pays for the AI?

People work with the brain in their own Claude, ChatGPT or Codex subscription. The chat in the app and meeting transcripts run on our server within your plan, so nobody needs another AI licence. Instead, the company can connect its own AI or its own key and pay for it directly.

Can we use our own AI?

Yes. The company can connect its own Anthropic key, a server with an OpenAI-compatible API (for example vLLM or Ollama), or its own Claude through a runner on its server. The runner only connects out from the server. An admin decides which feature runs on which AI.

Can onlybrain run entirely at our place?

Yes, that is Enterprise: onlybrain and the AI run on your server, with a local model, sign-in through Entra ID or Keycloak and in your look. If you don't want to run it yourselves, you can have a private instance with us in the Czech Republic. We are fine-tuning Enterprise with the first companies in a pilot and price it to fit you.

Does it work with Perplexity, Gemini or Copilot?

Today Claude, ChatGPT, Claude Code, Codex, Cursor, VS Code with GitHub Copilot and Gemini CLI work with the brain through the connector. Any tool on your computer can also open the brain as the folder the onlybrain app keeps. Perplexity, Gemini Enterprise, Microsoft 365 Copilot and Mistral Le Chat support custom connectors, and we are now checking the onlybrain sign-in with them.

How is our data protected?

The connection is always encrypted (HTTPS). The onlybrain team does not look into brain contents, only at your explicit request. What a role can't see is not in its people's copy at all.

What format is the brain in?

An ordinary folder of files (Markdown, PDF, spreadsheets) in Git. Whoever has it downloaded has their part on their computer and can open it with any tool.

Does it work with Codex too?

Yes. Instructions, agents and skills are set up in the brain so they work in both Claude Code and Codex.

What is local AI, and does data leave the company?

The model runs on a graphics card in a server in your company and you control it in onlybrain. With Enterprise everything stays inside the company. In the onlybrain cloud the brain stays with us and local AI jobs go to your card. When the card is down or can't keep up, the fallback can take them over, depending on the settings (Claude Haiku on our account, with a spending cap), and by default the fallback is on. Turn it off and, depending on the settings, the job waits or fails, and it never goes to a third-party AI. The server with the card only connects outward.

What hardware does local AI need?

A server with a graphics card with 32 GB of memory. We plan for AMD and NVIDIA cards. We measure on the AMD Radeon AI PRO R9700. The profile for the NVIDIA RTX 5090 is ready, but we have not yet verified it on a real card. Other cards get an estimate from their memory, labelled unverified, which the calculator refines.

Is local AI as good as Claude?

Not always. Open models on a single card are weaker than the largest cloud ones. They are enough for summaries and file descriptions. Meeting transcripts on the card are coming, and we only run chat and agents on a local model after measuring it on your own tasks. That's what Measure and tune in onlybrain is for.

Cloud or our own server?

You connect your card to the onlybrain cloud as a runner: the brain stays with us, the jobs run on your card. If you want everything to stay inside the company, there is Enterprise with onlybrain and the AI on your server. We are fine-tuning local AI with the first companies in a pilot and price it to fit you.

Give the whole company one brain

We're looking for the first companies for a pilot, in the cloud or inside your company. We'll help you build the brain.