Most AI agent tools hand the model a chat box and a few functions. Agent Zero hands it a computer. Not a metaphorical one — a Dockerized Linux desktop with a real browser, a terminal, code execution, and a file system the agent can actually live in. If the framing "give the agent a computer and get out of the way" appeals to you, this is the most complete open-source take on it, and it's free.

What it is#

Agent Zero is a self-hosted, open-source (MIT) multi-agent framework. Two words in that sentence do the work.

Self-hosted: you run it, usually in Docker, and drive it from a web UI on your own machine. Nothing is sent to a vendor's cloud except the LLM API calls you configure.

Multi-agent: it's hierarchical. A primary agent takes your instruction, breaks it into parts, and spawns subordinate agents to handle them — so "research these five competitors and draft a comparison" becomes a small crew, each agent browsing, running code, or writing, then reporting back up. The whole thing has a real shell, a browser with DOM annotation, and a bridge back to your host for files and commands.

Who it's for#

Founders and builders who want an open-ended operator, not a single-purpose tool. If your task is "change this codebase," a focused open-source coding agent like Aider, Cline, or OpenHands is tighter. Agent Zero earns its weight when the job spans a browser, a terminal, and a file system at once — scraping and reshaping data, driving a web app that has no API, running a multi-step automation, or prototyping an agent that needs a real computer to be useful.

The v2 line (current release v2.4, July 10, 2026; roughly 18k GitHub stars) is the one to evaluate. It matured the framework from a clever demo into something you can actually organize work in:

How to start#

One command:

docker run -p 80:80 -v a0_usr:/a0/usr agent0ai/agent-zero

Open the web UI, drop in a model API key in settings, and start a project. There are also one-line installers — curl -fsSL https://bash.agent-zero.ai | bash on macOS/Linux, irm https://ps.agent-zero.ai | iex on Windows.

It's model-agnostic through a LiteLLM-based config: OpenAI, Anthropic Claude, Google Gemini, xAI Grok, OpenRouter, Bedrock, or fully local models via Ollama. You set the chat, utility, and embedding models separately — a strong model to reason, a cheap or local one for the grunt calls.

What it costs#

The framework is free, MIT-licensed, self-hosted — no license fee, no required paid tier. Your only bill is the LLM API usage the agent generates. Point it at local Ollama models and that bill goes to zero, paid instead in the hardware to run them.

The trade Agent Zero asks you to make is control for power. You give an AI a real shell and a real browser; in return you own the blast radius. That's not a bug in the design — it's the design. Sandbox accordingly.

The honest catch#

An agent with a shell, a browser, and a host bridge executes arbitrary code — that capability is the whole point and the whole risk. Run it inside its Docker sandbox, scope any credentials to the minimum, never expose the web UI to the open internet, and vet plugins before installing (the 100+ in the hub are community-contributed and pass a policy-plus-review check, not a guarantee). It's also heavier than a library — a whole Linux environment per instance — and it moves fast: the jump from v1.20 to v2.4 happened in weeks and included breaking changes, so pin a version and read the release notes before you upgrade.

Used with those guardrails, Agent Zero is the clearest open-source answer to a question more founders are asking this year: what does an AI do when you stop handing it functions and start handing it a machine? Spin one up, give it a small real task, and watch the subordinate agents fan out. It's the most concrete way to feel where autonomous work is actually going — for the cost of a docker command.