Choose Agent Zero if you are comfortable with Docker and want an agent that works on a real Linux machine you own, with projects kept apart. It replaces a hosted chat assistant for hands-on computer tasks, but you supply the model account, and the built-in plugin hub lists more than 100 community add-ons.
What changed
Agent Zero is an open-source AI agent framework from agent0ai, released under the MIT licence. It runs in Docker and gives the model a Linux desktop, a terminal and a browser, and you connect a model of your choice. Its GitHub repository has passed 19,500 stars, and version 2.14 shipped on 13 October 2026.
Maker: agent0ai · Autonomy: semi autonomous · Maturity: GA
Underlying models: Bring your own (hosted providers, OpenAI Codex plan, local models)
About Agent Zero
Agent Zero is an open-source agent framework from the GitHub organisation agent0ai that runs inside Docker and gives the model a full Linux computer to work on. Its README calls it a framework for work that needs more than chat, and by October 2026 the repository had passed 19,500 stars and 3,800 forks. The project does not publish who runs it or where it is based, so this page names no founder or office. Where the OpenClaw family is built around chat-app reach, Agent Zero is built around the machine the agent sits in.
You open a web interface, connect a model provider, and the agent works inside a Dockerized XFCE desktop that appears in a Canvas beside the chat. It can run terminal commands, open desktop software, browse the web with a built-in browser, and edit Markdown, Writer, Spreadsheet and Presentation files in LibreOffice while you type in the same document. It supplies no model of its own: you connect a hosted provider (the v2.14 notes mention routes through OpenRouter, Z.AI and Nebius), your OpenAI Codex plan, or a local server such as Ollama, with LiteLLM handling the translation between them. Every agent can also create subordinate agents, so a researcher, a coder and a reviewer can each keep a focused context. Because it operates files, a browser and a shell instead of only replying, it counts as an agent rather than a chatbot.
Typical jobs are a scoped code change in a cloned repository, a spreadsheet model built together with you, a web page you annotate element by element so the agent can rebuild it in your stack, and recurring scheduled checks. Projects keep files, instructions, secrets, memory and model presets separate per piece of work, and a feature called Time Travel keeps snapshot history of the agent's workspace so you can diff and revert. An optional A0 CLI Connector lets the same agent reach files and a browser on your own machine instead of only the container. A hosted assistant answers inside a browser tab, while this one acts in an environment you own. Browse the other personal-agent records for contrast, starting with the hosted Manus, the self-hosted Hermes Agent and the container-per-session NanoClaw. Big-vendor takes appear in Gemini Spark and in ChatGPT Work. For lighter assistants, Dots is a place to start, as is xAI's Grok Bot.
There is no licence fee, because the code is MIT. You pay your model provider by usage, or only for hardware if you run a local model, and the README says it runs wherever Docker runs, including a small VPS or a Raspberry Pi. The recommended route is the A0 Launcher desktop app for macOS, Linux and Windows, there is a one-line terminal installer, and Docker users can start it with a single docker run command.
Release v2.14 arrived on 13 October 2026, following v2.13 on 23 September and v2.12 on 9 September, so the project ships roughly every two to three weeks. The v2.14 notes report about 27% faster startup to first token and a 99.3% prompt-cache hit rate in a live test, both vendor-reported. Treat it as fast-moving software: the README tells users to keep it in Docker, avoid mounting a whole home directory, and keep backups of important workspaces.
Pricing
No fee applies to Agent Zero itself. Your model provider bills you by usage, a local model server costs only your own hardware, and a rented VPS adds its own hosting fee. The closest relative with the same cost shape is [OpenClaw](/hub/agents/openclaw), and the lighter Python option is [nanobot](/hub/agents/nanobot).
Key Features
- Linux desktop in the Canvas: The agent drives a real XFCE desktop inside the container, so it can open terminals and GUI software such as Blender while you watch and take over.
- Browser with DOM annotation: A built-in browser lets you click a page element to inspect it, change it, lift it into your own stack or leave a review comment for the agent.
- Live document cowork: Markdown, Writer, Spreadsheet and Presentation files open in the Canvas so you and the agent edit the same document at once.
- Projects with isolated memory: Each project holds its own files, instructions, secrets, memory, Git repositories and model presets, which keeps client work apart.
- Plugin Hub and skills: A built-in hub lists more than 100 community plugins, and skills load on demand or stay pinned to a chat.
- Multi-agent delegation: Any agent can spawn subordinate agents that keep their own context and report back, and each child chat shows as working or paused.
- A0 CLI Connector: Install the a0 command on your own machine and the sandboxed agent can reach local repositories, a shell and your own Chrome, Edge or Brave browser.
- Time Travel snapshots: Workspace snapshots under /a0/usr let you diff what changed and revert, though the README says it does not replace Git or backups.
Strengths
- Free under an MIT licence with about 3,800 forks, so you can read, change and redistribute the code.
- A full Linux desktop, terminal and browser sit inside one Docker container, so work with GUI software needs no separate setup.
- You choose the model, from a hosted provider or the OpenAI Codex plan you already pay for to a local server, and can switch with named presets.
- Active development: three releases shipped between 9 September and 13 October 2026.
Weaknesses
- No model is included, so you need a provider account or local hardware, and the running cost depends on the service you pick.
- It needs Docker and the agent has a real shell, so the README warns against mounting your whole home directory and advises review of actions touching money or accounts.
- The project does not name its company, team or location on the pages read, so there is no vendor to hold to a support promise.
- Moves fast: it ships a release every few weeks, and no independent benchmark scores were found for it.
Frequently Asked Questions
What does it cost to run Agent Zero?
Nothing is charged for Agent Zero itself, so the real cost is the model: a hosted provider bills you by usage, while a local server such as Ollama costs only your hardware and electricity. A rented VPS adds its own monthly fee, and the README says the software runs wherever Docker runs, down to a small VPS or a Raspberry Pi. Check your provider's price list before leaving scheduled tasks running.
Will Agent Zero act without asking me?
It acts through the tools in its container, including a real shell, and you can watch every step in the Canvas and step in with your own mouse and keyboard. The README advises reviewing actions that touch accounts, money, production systems or private data. Giving the A0 CLI Connector read and write access to your own computer widens what it can change, so grant that only for machines and folders you trust.
Which self-hosted agents compete with Agent Zero?
OpenClaw is the best-known self-hosted family and fits people who want reach through chat apps. nanobot is a lighter Python gateway with a browser workbench, and NanoClaw wraps each session in its own container. Hosted choices such as Manus remove the server but move your data to the vendor.
How is Agent Zero different from nanobot?
Agent Zero centres on a Linux desktop, browser and document editor that the agent operates inside Docker, plus per-project memory and secrets. nanobot centres on a gateway that you reach from a browser, a terminal or chat apps and that works on folders you choose. Choose Agent Zero to watch an agent use software, and nanobot for a smaller footprint.
What does the first hour with Agent Zero look like?
Install Docker, then either download the A0 Launcher or run the one-line installer for your system, or use the docker run command from the README and open the web interface. Add a model provider in settings, then give it one concrete task such as building a small spreadsheet. Install the a0 command later if you want it to reach files outside the container.
Top Alternatives
- OpenClaw: Pick Agent Zero for a Linux desktop and browser inside Docker; OpenClaw is the closest named relative if chat-app reach matters more.
- nanobot: Agent Zero suits people who want the agent to operate a whole machine, while nanobot is a lighter Python gateway for browser, terminal and chat apps.
- NanoClaw: Agent Zero keeps one long-lived container with a desktop, whereas NanoClaw starts a fresh Docker container for each session.
- Hermes Agent: Pick Agent Zero for visible desktop and document work, and Hermes Agent if an agent that learns from your routines matters more.