Local Cognitive
On your computer
Models run through a built-in llama.cpp: Metal on Apple silicon, Vulkan or CUDA on Windows. Pick one from Hugging Face; memory is checked before the download.
Latest version · macOS · Windows · Linux server
Private AI on your own machines.
Chats, agents and workflows with models that run on your computer, or on your own GPU server that you connect with one command.
Set up a Linux GPU server in one commandA short animation: a chat moves from this computer to the GPU server cluster0, which answers with a model running on CUDA.
See it in action
Forty seconds of the real app.
Recorded on a Mac and a Linux GPU server. Where waiting is sped up, the video says so.
On your GPU server
One command installs it on Linux, and a one-time key connects your computer from anywhere. Chats, models, workflows and Synthesis run there too.
Private by design
Your computer and your server talk end to end encrypted. The cloud sees accounts and token counts, never prompts, answers or files.
Workflows
Agents you can watch, step by step.
Chain agents, files, commands and MCP tools on a canvas. Run it as a task or on a schedule, follow every step live, and approve the ones that matter.
Workflows guideA short animation: a workflow runs an agent, saves its findings, runs the tests and waits for an approval.
A short animation: Synthesis writes a rate limiter, two of six checks fail, the second iteration passes all six and is accepted.
Synthesis
Code that has to pass its contract.
Describe a module and its checks, then let agents write it: one local model, or tens of agents in parallel on local and cloud models, each reading and writing only the files you allow. Failed checks go back to the agents until everything passes, and every iteration stays on record.
Models
Any model on the hardware you have.
Search Hugging Face's GGUF models and download any that isn't gated. Before a download, Local Cognitive estimates the memory and disk a model needs and tells you whether it fits. Downloads pause, resume and are checked by SHA-256, and loading one model never unloads another.
- Split a model too large for one GPU across several, by layers or, with CUDA, by rows.
- Run different models on different GPUs at the same time.
- Models made of several GGUF files download and install as one.
- Import a GGUF you already have, with all its parts.
Several GPUs work on Windows and on a Linux server. They are not yet tested on real multi-GPU hardware.
Models guide| Where | How models run |
|---|---|
| macOS, Apple silicon | Metal, sharing the Mac's unified memory |
| Windows | Vulkan on any dedicated GPU; CUDA on NVIDIA, a one-time download from the app |
| Linux server | CUDA on NVIDIA, from Maxwell to Blackwell; the CPU without a GPU |
Built on llama.cpp. Nothing else to install.
Your GPU server
One command on the server. One key in the app.
Install the server on a Linux machine with NVIDIA GPUs and use it from wherever you are. Work keeps running while your computer sleeps.
Server guide# on your Linux server
$ curl -fsSL https://github.com/…/install.sh | sudo bash
$ sudo local-cognitive-server pair
# prints a one-time key: LCR1-…
# then in the app: Remote → Connect a server
Your computerLocal Cognitive for macOS or Windows
RelayEncrypted bytes only
cluster0Your GPU server
TLS 1.3 runs between your computer and your server, and each checks the other's key. Both only connect out, so you open no ports and need no domain or certificate.
- 01Download the app
For a Mac with Apple silicon or a Windows PC. Pick a model and chat, no account needed.
- 02Install the server
One command on Linux with systemd. With NVIDIA driver 570 or newer, it adds CUDA once you accept NVIDIA's license.
- 03Connect with a key
Sign in, then paste the key in Remote → Connect a server. The first account to connect owns it.
And everything around it
Projects
A project is a folder on your disk with its own chats. Agents read, search, edit and run checks inside it.
Review edits
Open a file an agent changed, comment on a selection and send it back. The agent changes only that part.
Approvals
Each chat asks first, approves for you, or has full access. An approval shows exactly which file or command it is for.
MCP servers
Connect Blender, your editor or other MCP servers, local or over HTTP. Import the ones from Codex, Claude Desktop or Cursor.
Plugins
Notion, Linear and Jira, read only or read and write. Every write asks first.
Subagents and hypotheses
Call agents with @Name, or test an idea in Hypothesis mode: one agent supports it, one attacks it, and a judge weighs the evidence.
Images
Attach images for a vision model. A model that can't see them says so instead of guessing.
Voice input
Dictate with whisper.cpp on your own computer. The audio never leaves it.
Cloud models, when you want
Add OpenAI, Anthropic or Gemini beside local models, and mix them in one chat.
Privacy
Your work stays on your machines.
Local chats work without an account. Sign in only to use Remote and to see usage across your computers.
Keys and tokens are stored encrypted on your devices. Error reports go to Sentry in the EU, and only if you turn them on.
How your data is kept- The Cloud sees
- Your account
- Your servers and computers
- Which computer connects to which server, and when
- How much data passes
- Token counts for Usage
- It never sees
- Prompts and answers
- Your files
- Model output
- Your provider keys
Get Local Cognitive
Latest version · Release notes
macOS
Apple silicon · macOS 13.3 or later
New versions install from the app in a few clicks.
Download .dmgWindows
Windows 10 and 11 · x64
The installer isn't code-signed yet: choose More info → Run anyway.
Download .exeLinux server
x64 with systemd: Ubuntu 22.04+, Debian 12+, Fedora 36+ or Amazon Linux 2023. NVIDIA optional.
curl -fsSL https://github.com/Ilyaberdar/local-cognitive-AI-system/releases/latest/download/install.sh | sudo bash
Documentation
Read the docs
Step-by-step guides for the app, your server and Synthesis.
Get startedInstall the app, pick a model and start a chat.
ModelsThe catalog, memory checks, imports and GPUs.
RemoteHow your computer and your server connect, and what each one keeps.
GPU serverInstall, pair, update and choose GPUs.
WorkflowsSteps, tasks, schedules and approvals.
SynthesisContracts, agent flows and the language.
Questions
Do I need an account?
Not for local chats and models. Sign in to use Remote with your own server and to see token usage across your computers.
Does anything leave my computer?
Not with local models. If you choose a cloud model, that chat goes to its provider, and traffic to your own server is encrypted end to end. Error reports go to Sentry only if you turn them on.
Which GPUs work?
Apple silicon through Metal. On Windows, dedicated GPUs through Vulkan, and NVIDIA through CUDA. On a Linux server, NVIDIA with driver 570 or newer; without a GPU, models run on the CPU.
What does a server need?
Linux x64 with systemd (Ubuntu 22.04+, Debian 12+, Fedora 36+ or Amazon Linux 2023), outbound HTTPS, and about 1.5 GB of disk plus your models. No open ports, domain or certificate.
How do updates work?
The app tells you when a version is out. On a Mac it downloads and installs it in a few clicks; on Windows it links to the new installer. A server updates with sudo local-cognitive-server update and rolls back by itself if the new version fails its health check.