Cortex

Cortex is the MindStone Agent command center: a self-hostable web UI for your local AI stack.

It gives a local AI workstation or lab box a single landing page and live dashboard for:

system and GPU health
host OS, kernel, and NVIDIA driver versions
live performance telemetry
installed local Ollama model inventory
per-model context windows
model discovery across the Ollama library and Hugging Face
local Ollama pulls with live progress
configurable local service catalog
health checks for tools such as Ollama, Open WebUI, ComfyUI, Jupyter Lab, Synapse, and more
runtime theming and theme selection
opt-in local tool management

Repository:

https://github.com/MindStone-Agent/cortex

Status

Cortex is public and in active development.

The repository is now tagged v0.5.0 on main at commit 87162d9. It is usable for trusted local environments, but should not be treated as a hardened public internet service.

Target environment

Cortex was built specifically for NVIDIA GB10-based platforms and developed on the NVIDIA DGX Spark reference platform.

It can be useful on other Linux hosts with local AI tooling, but the GB10/DGX Spark path is the primary design target.

Key features

Dashboard — CPU, GPU, disk, uptime, host versions, and currently resident Ollama models.
Live performance stream — Server-Sent Events telemetry at roughly 1 Hz.
GB10-aware GPU signal — on unified-memory GB10 systems, GPU power draw is treated as the reliable activity signal because nvidia-smi utilization can remain high while a process is merely resident.
Models on the dashboard — installed local Ollama inventory, including resident models and context-window information.
Model discovery + pull — search the Ollama library and Hugging Face, manually refresh for same-day model drops, and pull selected models to local Ollama with live progress. This is discovery and local pull orchestration, not remote hosted-model execution.
Services catalog — configurable service cards with health checks and direct links.
Status row — quick status visibility for the local AI machine and configured services.
Auto-discovery — probes localhost for common AI-tool ports through /api/discover.
Settings panel — toggle branding/logos and manage supported tools from the UI.
Opt-in tool installs — Docker-based tools can be installed from Settings when explicitly enabled.
Opt-in Ollama update/restart — available only after installing a scoped root-owned wrapper and enabling the config flag.
Opt-in Tailscale control — available only after installing a scoped wrapper and enabling the config flag.
Runtime theming and theme selector — theme.json controls logo, subtitle, palette, and theme presets without rebuilding.

Quick start

One-command install on a fresh GB10 / Ubuntu box:

curl -fsSL https://raw.githubusercontent.com/MindStone-Agent/cortex/main/scripts/install.sh | bash

Manual development path:

git clone https://github.com/MindStone-Agent/cortex.git
cd cortex
pnpm install
cp cortex-config.example.json cortex-config.json
pnpm dev

Development server:

http://localhost:3000

Production is typically:

pnpm build
pnpm start

usually behind Caddy.

Configuration

Cortex uses runtime configuration files so local setup does not need to be committed:

cortex-config.json — local services, ports, health paths, icons, and system action flags.
theme.json — branding, logo, subtitle, palette, and theme presets.

Models on the dashboard are read from a local Ollama instance:

localhost:11434

The model discovery page searches upstream catalogs and can pull selected models into the same local Ollama instance with live progress. Discovery refresh is manual so same-day model drops can be picked up without waiting for a site deploy.

Configuration should keep local machine details in runtime files, not source control.

Security model

Cortex is intended for a single trusted machine or trusted private LAN.

By design, it currently has:

no built-in authentication
plain HTTP by default
optional privileged actions only when explicitly enabled

Do not expose Cortex directly to the public internet or an untrusted network as-is. Put authentication in front of it first, keep privileged actions disabled unless you need them, and prefer Tailscale, WireGuard, or SSH tunnels for remote access.

The repository includes a dedicated security document:

SECURITY.md

Relationship to MindStone Agent

Cortex is not the MindStone Agent continuity substrate itself. It is the local command center for the machine running the local AI stack.

It complements the rest of the MindStone Agent product family:

MindStone for Claude Code and MindStone for Pi provide continuity layers.
MindStone Agent Harness is the coming standalone agent runtime.
Synapse provides agent-to-agent and human review channels.
Cortex gives the local AI machine an operational dashboard and trusted-LAN command surface.