If you are new to Artificial Intelligence, you might be used to websites like ChatGPT or Claude. Those are "Cloud AI" systems. They live on someone else's computer, they read your private data, and they restrict what you can ask them. OpenZero is fundamentally different.
OpenZero is a Sovereign AI Agent. An AI Agent is a software system that doesn't just talk to you; it can observe its environment, make plans, and execute tasks autonomously. OpenZero is installed directly onto your hardware. It acts as your digital employee, complete with terminal access to your computer, a web browser (Moltbot) to see the internet, and a direct uplink to your phone via Telegram.
Deploying the OpenZero Lattice is designed to be fully automated. The installer downloads the Neural Brain (Python), the Vision System (Node.js/Chrome), and the Process Manager (PM2), linking them all together.
To install OpenZero on a new Linux server, simply run this command in your terminal:
If you need OpenZero to work on a machine that will not touch the internet after setup, use the offline release path. Build the bundle on one connected Linux OpenZero node, then move the finished archive to the air-gapped target. In practice the archive is usually 15 GB to 25 GB because it carries the OpenZero code, Python wheels, a local Node runtime, PM2, an Ollama runtime, bundled Moltbot dependencies, and the local Gemma model store from the builder node.
The builder writes dist/openzero_offline_release.tar.gz plus a matching SHA-256 file.
/build-offline serves the builder script, /install-offline serves the offline installer, and /offline-guide serves the beginner written guide. The offline installer now treats the extracted folder as the final install directory by default, so it does not waste disk space duplicating a 15 GB to 25 GB bundle unless you explicitly pass a different --dir path. If you upload a finished bundle to the web root using the exact filename openzero_offline_release.tar.gz, the stable public download URL becomes https://openzero.talktoai.org/openzero_offline_release.tar.gz.OpenZero receives constant upgrades. You can safely run the update command at any time. The updater is designed to fetch the newest code, UI patches, Ollama runtime refreshes, and local model repairs without deleting your private .env keys or your chat logs.
STOP control. It does not just hide the spinner; it sends a stop request so Agent Zero halts at the next safe boundary in the operator loop. If Hive backlog is distracting you, use PAUSE HIVE or CLEAR QUEUE; that queue is only unsent lattice events, not your local LLM task list.OpenZero keeps Gemma via Ollama as the default local lane. If you want the optional Microsoft BitNet CPU-efficient path for lower-power or older systems, install it as an add-on instead of replacing the whole runtime stack.
bitnet.cpp path and the official GGUF weights, with a maximum context window of 4096. If you do nothing, OpenZero stays on the default Gemma/Ollama lane.While OpenZero runs perfectly on AlmaLinux, Ubuntu, or Debian cloud servers, Linux Mint is the absolute best environment for this framework. By installing OpenZero on a physical or RDP-accessible Linux Mint machine, you get a full graphical desktop environment while simultaneously running your Sovereign AI Node silently in the background. It is a desktop PC and an autonomous AI Server built into one seamless machine.
If you are brand new, think of OpenZero in three layers: your node, your Hive HQ, and optional federation mirrors. Your node is your own machine. Hive HQ is the main coordination hub you point to. Federation mirrors are backup hubs that can keep the network alive if one operator or one database disappears.
| Setting | What it means in plain English |
|---|---|
HIVE_MIND_ENABLED=true |
Your node participates in the Hive and registers itself instead of staying fully private/local-only. |
OPENZERO_HIVE_MODE=standalone |
Use one Hive HQ only. This is the simplest mode and best for a first install. |
OPENZERO_HIVE_MODE=federated |
Use one Hive HQ plus one or more backup mirror URLs. OpenZero writes to all reachable mirrors and queues failed writes locally for replay later. |
OPENZERO_HIVE_MODE=local |
Stay operational with local continuity data only. This is the emergency mode when all remote Hive endpoints are gone. |
OPENZERO_HIVE_URL |
Your primary Hive HQ address. Example: https://openzero.talktoai.org/api/hive. |
OPENZERO_HIVE_MIRRORS |
Comma-separated backup Hive endpoints. Example: https://hq2.example.org/api/hive,https://hq3.example.org/api/hive. |
OPENZERO_HIVE_LOCAL_SPOOL_ENABLED=true |
Keeps a local queue of node events and knowledge updates so they can be replayed when a Hive endpoint comes back. |
OPENZERO_HIVE_BACKGROUND_PUSH=true |
Sends lattice contributions in the background so Hive sync does not hold up the main reply path. |
http://localhost:1024 in your browser.STANDALONE for your first setup.FEDERATED mode and add mirror URLs.OpenZero runs continuously in the background so it is always ready to receive commands from your web dashboard or Telegram. It uses a professional process manager called PM2 to stay alive. If you ever change a setting, or if the AI gets stuck thinking, you do not need to restart your computer. Just use these PM2 commands in your terminal:
| Command | What it does |
|---|---|
pm2 restart all |
The Master Reset. This fixes 99% of all problems by restarting both the Brain and the Vision systems simultaneously. |
pm2 status |
Shows a table of your running systems. Use this to check if OpenZero is online and how much CPU/RAM it is consuming. |
pm2 logs |
Shows a live matrix feed of what the AI is thinking behind the scenes. Excellent for debugging if a command fails. |
pm2 restart zero-brain |
Restarts only the logic/chat engine. |
pm2 restart zero-vision |
Restarts only the Moltbot headless browser (useful if website screenshots start failing). |
When you access OpenZero via http://localhost:1024 (or your server's IP address on port 1024), you are greeted by the Super Panel. This is your command center.
/resolve/ link), choose a model alias, and click "Pull Weights" to inject the new model into your local system. Custom GGUF packages are now tracked in their own inventory so you can delete the stored file and linked Ollama alias cleanly later.OpenZero operates on a Hybrid Neural Lattice. This means you have the power to choose where the AI's "brain" lives, instantly toggling between them in the System Config menu.
Fresh OpenZero installs now default to the Gemma 4 edge track through Ollama. The normal default is gemma4:e4b, while lower-RAM boxes can use gemma4:e2b. Stronger nodes can move up to gemma4:26b or gemma4:31b. This means the AI processes your questions on your own machine using your own CPU/GPU.
Benefits: Zero cost, maximum privacy, and air-gapped capability. Sensitive information never leaves your environment.
./models folder. Ollama keeps its own model store. The local ./models folder is reserved for custom GGUF files that you download and inject manually.gemma3:4b and gemma3:12b remain compatibility fallbacks for older runtimes.For highly complex math, deep coding problems, or massive context windows, you can enter a Groq API Key in the settings. Groq uses specialized supercomputer chips called LPUs to process AI responses at lightning speed.
The Groq Compound Model: We highly recommend the groq/compound-large model for OpenZero. It is engineered specifically for autonomous agent workflows and recursive, multi-step planning.
Agent Zero isn't just a chatbot; it is a digital entity with tools. By simply typing natural language, the AI will autonomously deploy specialized XML tags to take action in the real world.
| Capability | How it Works |
|---|---|
| Terminal Sovereignty | The AI has direct Bash access. If you type "check my active ports", the AI will secretly write <bash>netstat -tulnp</bash>, the system will execute it on your server, and the AI will read the results back to you. |
| Structured File & Code Ops | Agent Zero now has a stronger local operator lane for code and project work. It can read files with line ranges, write or append files, replace text, build directory trees, search source trees, create folders, and remove paths with explicit local operator actions before it falls back to raw shell. |
| Archive Operations | OpenZero can inspect ZIP archives, create new bundles, or extract them into a target folder. This makes it easier to unpack model bundles, backup project folders, or ship artifacts without forcing the operator to type archive commands manually. |
| OSINT (Digital Recon) | OpenZero acts as a cyber-investigator. If you command it to "run recon on John Doe", it will deploy an <osint> tag. The system uses advanced "Google Dorking" parameters to bypass search noise and pull raw footprints from LinkedIn, Twitter, and exposed PDF files directly into the chat. |
| Satellite & Moltbot Vision | OpenZero has eyes. Moltbot is a hidden Google Chrome browser living inside the server. If you command the AI to "go to talktoai.org", Moltbot will silently load the page, take a screenshot, display it in your sidebar, and read the text. You can even type "show me satellite imagery of Nottingham, UK", and the AI will lock the coordinates and pull the visual feed. |
| Web Fetch & Search | Besides Moltbot browsing, OpenZero can now pull a webpage into readable text directly or run a Serper-backed web search from inside the agent loop. This is useful when you want the node to read docs, fetch a release page, or gather search results quickly without a full visual browser pass. |
| Remote Node Ops | When SSH tooling is present, Agent Zero can execute key-based SSH commands and SCP transfers against other boxes. That gives OpenZero a clean remote-ops lane for fleet work without forcing every task through a raw one-line Bash answer. |
You can securely talk to your OpenZero node from anywhere in the world using your smartphone via Telegram. When you message your bot, the text routes directly to your home server, OpenZero processes it, and texts you back.
/newbot and hit send.openzero_alpha_bot).http://localhost:1024, use the left-side control panel, paste the token into the Telegram field, and hit Save. Your personal bot is now online and listening to your commands.A tactical list of commands you can type directly into the Agent's chat box to manage your Linux infrastructure. Remember, you can prefix any standard Linux command with the word run to force the AI to execute it.
run curl ifconfig.me (Check Public IP)run netstat -tulnp (View Active Ports)run ping -c 4 google.com (Check Latency)run ss -tupln (Socket Statistics)
run htop -b -n 1 (CPU/RAM Snapshot)run df -h (Check Disk Space)run free -m (View Memory Blocks)run uptime (System Load Average)
run docker ps -a (List All Containers)run docker logs [id] (View Container Logs)run systemctl status nginx (Service Check)run git pull (Update Code Repositories)
run tail -n 50 app.log (Read System Logs)run zip -r backup.zip . (Create Data Archive)run chmod +x script.sh (Make File Executable)run ls -la (List Directory Contents)
If you downloaded the ZeroMint OS ISO, you possess a complete, bootable operating system with OpenZero permanently baked into its core. To deploy this to physical hardware:
Use these exact credentials to log into the ZeroMint desktop after booting:
zero1234ZERO
SECURITY OVERRIDE PROTOCOL: If you change the default password using the standard Linux `passwd` command, the AI agent will lose root access and fail system tasks. You MUST change your password using the internal Agent Sync utility:
This script securely changes your Linux Desktop password AND injects it into the Agent's memory so it retains Terminal Sovereignty.
The Super Panel allows you to dynamically download new "brains" for your AI using GGUF files. GGUF is a specialized file format that compresses massive neural networks so they can run efficiently on standard consumer hardware.
If you want the official Google local path, use the built-in Gemma 4 install buttons first. Use Hugging Face injection only when you have a direct GGUF file link and want a custom model alias inside OpenZero.
When browsing Hugging Face (like huggingface.co/shafire), you will often see different "Quantization" levels like Q4, Q6, or Q8. These represent the compression ratio:
Operating a Sovereign Node requires vigilance. If you open your server to the world, bad actors will attempt to scan it. We highly recommend configuring the Uncomplicated Firewall (UFW) to lock down your node.
Q: Why does the AI say "[ERROR] Local Brain Offline"?
sudo systemctl restart ollama, then wait 30 seconds and verify with curl http://127.0.0.1:11434/api/tags. If you intentionally selected BitNet, use the panel controls for Install BitNet or Repair BitNet, or rerun curl -sL https://openzero.talktoai.org/update.sh | bash -s -- --bitnet.Q: Moltbot Vision isn't loading website screenshots.
pm2 restart zero-vision to reset the buffer. Also ensure your server has at least 2GB of free RAM, as Chrome will crash if memory runs out.Q: I changed my API key but the AI is still using the old one.
.env file, but Python loads these variables into memory at startup. You must restart the brain to flush the memory. Run pm2 restart zero-brain.Q: Can OpenZero run with no internet at all after I prepare it?
openzero_offline_release.tar.gz to the target machine, extract it, and run ./install_offline.sh. That path carries the code, bundled dependencies, Ollama runtime, and Gemma model store from the builder node so the target can stay air-gapped after deployment.Q: Can OpenZero really hack satellites?