Agent Zero

Agent Zero is an open source agent framework that gives the agent a full Linux system inside a Docker container: a terminal, a filesystem, a browser and a real desktop. It is not built for one task: it runs on its own computer, uses and creates tools, and keeps its work transparent.

The emphasis is on the environment: the agent treats the operating system as its tool, and where no built-in tool fits it writes code and runs it inside the container.

What it is

One container ships a full Linux system with an XFCE desktop session in the Canvas of the web UI, the same desktop the user sees.

• Desktop and browser. It can open GUI applications, manage files, read pages and fill forms. In annotate mode a click on a page element becomes an instruction.
• Extensibility. The Plugin Hub installs community plugins, while tools/, prompts/ and Skills take custom tools, instructions and procedures.
• Memory and projects. Projects separate files, instructions, secrets and memories per workspace.
• Cooperation. It acts as MCP client and server, A2A links several instances, and bounded work goes to subagents.

Install and run

Docker is the common path. On a desktop, A0 Launcher checks Docker, creates and manages instances and assigns ports; on a server, A0 Install can create one instance in quick-start mode and exit. With Docker running, the official image starts directly, with port 80 mapped and persistent data bound to /a0/usr. That directory holds files, chats and secrets, so use one clean mount and back it up.

Host access comes from the A0 CLI Connector, installed on the workstation rather than in the container. It connects the running agent to the local terminal, project files and browser while the agent stays sandboxed. Access is not automatic: host writes and command execution are granted per trusted workspace.

Models and providers

Three model roles exist: the main model talks and plans, the utility model handles quick internal work, and the embedding model powers memory and indexing.

• Cloud. OpenRouter by default, plus Anthropic, OpenAI and other hosted providers with your own key.
• Your account. Codex/ChatGPT, GitHub Copilot, Google Cloud Gemini or xAI Grok sign-in.
• Local. Ollama or LM Studio on the workstation, for offline use.

With local models the usual failure is the address: inside the container localhost points at the container, so a host model server is reached as host.docker.internal:11434. A server listening only on 127.0.0.1 cannot be reached: bind it to a Docker-reachable address and keep the port limited to trusted clients.

Documented limits

By the project’s own guidance, the power comes from a real environment and requires discipline.

• Isolation. Keep it in Docker or another isolated environment; mounting your whole home directory, or all of /a0 during an upgrade, is discouraged.
• Host access. Grant A0 CLI write and execution rights only on trusted machines.
• Secrets. Credentials live in /a0/usr/secrets.env, which is not always part of backup archives.
• Network. Never expose the web UI without authentication.
• Dependencies. Code execution needs Docker running, and chats and memory under /a0/usr can be lost without a backup or volume.
• Delegation. Subagents help when work splits cleanly; letting several agents edit the same files is advised against.

At CyberElectro we use Agent Zero when a task needs a real machine environment and long autonomous execution, next to Hermes Agent.

Further reading

• agent0ai/agent-zero - the source and the documentation tree, with prompts, tools and the safety model in it.
• Installation - the install paths through Launcher, A0 Install or Docker, plus local model setup.
• A0 CLI Connector - host access while the agent stays in its container.