Agent Zero
A self-hosted agent workbench for operating code, browsers, documents, and GUI software inside an isolated Linux desktop.
- Source repo
- agent0ai/agent-zero
- Stars
- ★ 20k
- Last updated
- 5d ago
- License
- NOASSERTION
- Primary language
- Python
- FA score
- 70/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A software engineer can inspect a repository, make scoped edits, run tests, and explain tradeoffs inside an isolated workspace with recoverable history.
- Not a fit if
- Teams unwilling to run Docker or keep a container running
- Organizations that cannot accept agent read/write and remote code execution on real hosts
- Source review
- 70/100 · Some gaps
What does this agent do, and when should you use it?
Agent Zero is an open agent framework for work that extends beyond a chat interface. It runs an XFCE Linux desktop, terminal, filesystem, built-in browser, and observable Canvas inside Docker, while allowing the user to intervene with a shared mouse and keyboard. Its browser can click, type, upload files, capture screenshots, and turn annotated DOM elements into inspect, change, lift, or review instructions. A live Markdown editor and integrations with LibreOffice Writer, Calc, and Impress support collaborative production of Markdown, ODT, ODS, and ODP files. Projects isolate files, instructions, secrets, memories, repositories, and model presets, while skills, plugins, MCP, A2A, custom prompts, and tools extend the framework. It is best suited to adopters who want a visible, configurable execution environment and are prepared to operate Docker and carefully control host access.
A task submitted through the Web UI can drive terminals, files, and GUI applications in the containerized XFCE desktop or use the built-in Browser to open pages, read them, click, type, upload files, and take screenshots. Browser Annotate mode can collect an element's DOM, styles, parent chain, and framework hints; apply and verify JavaScript changes; capture a component for reimplementation; or read review comments pinned to page elements. The Canvas Markdown editor creates and updates text files live, while LibreOffice Writer, Calc, and Impress let the user and agent work on ODT, ODS, and ODP files. Projects separate workspaces, memories, secrets, instructions, repositories, and Model Presets; Agent Profiles alter the current chat's working style, and Skills can be loaded on demand or pinned. A superior agent can delegate research, coding, analysis, or review to subordinate agents. With the A0 CLI Connector installed on a host, the same running instance can receive authorized access to real host terminals and repositories; Time Travel adds snapshots, diff inspection, historical state viewing, and reversion for Agent Zero-owned /a0/usr workspaces.
- A software engineer can inspect a repository, make scoped edits, run tests, and explain tradeoffs inside an isolated workspace with recoverable history.
- A frontend team reviewing a local web application can pin comments to specific DOM elements and ask the agent to implement the resulting UI fixes.
- A developer using a reference site for inspiration can capture a card or hero section with Annotate mode and reimplement it in the project's React and Tailwind stack.
- An analyst who needs an editable deliverable can ask for an ODS budget model with assumptions, monthly projections, and charts rather than receiving a static chat response.
- A creator who needs GUI automation can watch the agent operate Blender or another desktop application that lacks an API, then intervene through the shared desktop.
- A consultancy handling multiple clients can use separate Projects to isolate each client's files, secrets, memories, instructions, repositories, and model choices.
How do you install or deploy this agent?
For a personal machine, the recommended route is A0 Launcher v1.5, which checks Docker, creates instances, manages ports, and connects to local or remote Agent Zero installations. If Docker is already installed, run:
docker run -p 80:80 -v a0_usr:/a0/usr agent0ai/agent-zeroOpen the Web UI, configure an LLM provider, and submit a concrete task. The terminal installer for macOS or Linux is:
curl -fsSL https://bash.agent-zero.ai | bashOn Windows PowerShell, use:
irm https://ps.agent-zero.ai | iexFor a server or scripted setup, run:
curl -fsSL https://bash.agent-zero.ai | bash -s -- --quick-start --name agent-zero --port 5080If port 80 is occupied, use docker run -p 5080:80 -v a0_usr:/a0/usr agent0ai/agent-zero and open http://localhost:5080. Docker must be operational. Accounts or credentials needed for model access depend on the LLM provider selected in the Web UI.
How do you use this agent?
After opening the Web UI, configure an LLM provider and start with a specific, verifiable assignment such as inspecting a project, opening a webpage, or creating an ODS workbook. For browser collaboration, open the target page in Browser, enable Annotate mode, and select change, inspect, lift, or comment on an element; the agent can then act on the captured instruction. For durable documents, use the Canvas Markdown editor or create Writer, Spreadsheet, and Presentation files from the Desktop toolbar so both parties can continue editing them. Create a separate Project when files, repositories, secrets, memories, or model choices must remain isolated. To reach host files, install A0 CLI Connector on the host rather than inside the container: run curl -LsSf https://cli.agent-zero.ai/install.sh | sh on macOS/Linux or irm https://cli.agent-zero.ai/install.ps1 | iex in Windows PowerShell, then run a0 to connect to an existing instance. Grant host read/write or remote-execution access only to machines and workspaces you trust.
What are this agent's strengths and limitations?
- The container includes a real XFCE Linux desktop, allowing the agent to operate terminals, files, and GUI applications such as Blender while the user watches or takes over.
- DOM annotation converts specific webpage elements into inspect, change, lift, and review directives, providing more precise context than screenshots or prose alone.
- The Markdown and LibreOffice integrations produce editable Markdown, ODT, ODS, and ODP artifacts instead of trapping deliverables in chat history.
- Projects isolate files, secrets, memories, repositories, instructions, and model presets for parallel or client-specific work.
- Plugin Hub, Skills, MCP, A2A, custom tools, and custom prompts offer several extension layers, with prompts, tools, plugins, and settings remaining inspectable.
- Time Travel supplies snapshot, diff, and revert operations for agent-managed files under /a0/usr.
- The documented deployment paths depend on Docker; a stopped runtime or occupied port prevents the default startup flow.
- A0 CLI read/write access and remote code execution extend the agent's reach to real host files, creating material risk if permissions are granted too broadly.
- Browser activity, model providers, plugins, and remote installations can require network connectivity, provider accounts, credentials, and additional spending; no unified cost is documented.
- Time Travel covers Agent Zero-owned /a0/usr workspaces and is explicitly not a substitute for Git or backups.
- The repository license is reported as NOASSERTION, and the supplied material provides no license terms; adopters must verify rights to use, modify, and distribute it.
- The material explicitly documents OAuth for an OpenAI Codex plan, but it does not fully specify the availability or setup of other model providers; Gemini CLI and Claude Code are described only as future extra-usage integrations.
How does this agent compare with similar options?
The project describes Space Agent as a related, more polished product direction for an agent-shaped workspace. Agent Zero remains positioned as the open framework and Linux-powered workbench. The supplied material does not provide a feature-by-feature or compatibility comparison.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Agent Zero This agent | 70 · Some gaps | Desktop appFree + model costs | ★ 20k | 5d ago | Python | Codex |
| DeterminFlow | 66 · Some gaps | Desktop appFree + model costs | ★ 920 | 1mo ago | Python | — |
| ConnectOnion AI Agent Framework | 47 · Major gaps | CLIFree + model costs | ★ 1.5k | today | Python | OpenAI API · Claude API |
| ZhikunCode | 56 · Major gaps | Self-hosted serviceFree + model costs | ★ 525 | 1d ago | Java | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The README recommends Docker isolation, warns against mounting an entire home directory, and limits host read/write and remote execution grants to trusted machines and workspaces. It also describes project isolation, project secrets, and Time Travel recovery. Deductions apply because these are chiefly operator instructions: the supplied material does not show enforced confirmation gates for account, financial, production, or host actions, and it does not fully map data flows among model providers, browsers, plugins, MCP, and A2A. Requirements include explicit CVE remediation floors and many exact pins, but also broad minimum versions, a duplicate entry, no hashes, and no evidence of automated vulnerability scanning. The copyright holder, contact, and repository are identified, while per-plugin and external-component attribution is not shown.
The README's project, plugin, naming, and recovery descriptions do not visibly conflict with the supplied tests and release workflows. Tests cover migration idempotence, context budgeting, authentication/CSRF defaults, failure notification, and state updates. The deduction is that these tests concern only a small part of the framework and cannot establish repository-wide consistency. Dependencies are mostly constrained, Docker images are published for two architectures, and common installation failures are documented. However, operation still depends on Docker, model providers, remote installer endpoints, and a large third-party dependency set without demonstrated offline degradation or comprehensive availability controls. Error messages are concrete in tests and CI scripts, but no product-wide error taxonomy is evidenced.
The source clearly addresses software engineering, document collaboration, browser automation, desktop software, financial analysis, QA, servers, and host development. Multi-platform and multi-architecture installation plus projects, presets, skills, plugins, MCP, and A2A provide strong environmental adaptability. The safety model and container/host-bridge descriptions establish some boundaries, but permissions for individual tools, plugins, and subagents are not comprehensively specified. Skills may be loaded on demand or pinned and tasks are prompt-driven, yet the supplied files do not demonstrate conflict resolution, false-trigger prevention, or deterministic routing, so trigger precision remains thin.
The README has strong navigation, quick starts, deep-dive guides, a task-oriented documentation index, troubleshooting, and numerous examples. Launcher, scripted, server, and direct-Docker installation paths are concrete and usable. Naming is generally stable, although Agent Zero, A0 Launcher, A0 Install, A0 CLI, and the related Space Agent create several surfaces whose boundaries require interpretation. Docker, host-mount, backup, and Time Travel limitations are stated, but there is no comprehensive known-limitations register. The LICENSE is an explicit MIT license with holder, contact, and warranty disclaimer, justifying full credit despite NOASSERTION external metadata. Release automation supports branches, version tags, and GitHub Releases, but no standalone changelog is supplied. Organization, email, Issues, and community channels provide an update path, while the publisher remains unverified and no named maintainer or response commitment is evidenced.
The framework targets reusable artifacts—editable Markdown, ODF documents, code, browser changes, and desktop files—and adds live coworking, project isolation, extensions, and recoverable history. Those features show substantial practical usability and marginal value beyond a chat-only agent. Cost-benefit is deducted because the low-cost VPS claim and model presets are not accompanied by resource benchmarks, representative model spend, latency measurements, maintenance burden, or scaled-cost comparisons.
Major feature claims point to in-repository guides, while the supplied tests trace selected claims about migration, naming limits, failure notification, authentication/CSRF, and scheduler updates. The Docker workflow provides a second source type for versioned, multi-architecture publishing. Deductions apply because many core capabilities and the “100+” plugin claim appear only as README assertions and are not corroborated by the supplied implementation, inventory, or tests. The prose also does not consistently distinguish demonstrated facts, intended design, illustrative scenarios, and future integrations.
- This is a low-confidence static review of only the supplied files; no code, image, installer, or test was executed.
- The agent can receive a full Linux environment, browser control, host filesystem access, and remote execution capability. Preserve container isolation, minimize mounts, and impose external confirmation for actions involving accounts, money, production systems, or private data.
- The README offers installation by directly executing remote curl or PowerShell content. Pin and inspect those scripts and their provenance before deployment.
- Community plugins, MCP, A2A, and external model providers expand both supply-chain and data-disclosure surfaces; the supplied evidence lacks a complete data-flow map, plugin-review policy, or permission inventory.
- Time Travel covers Agent Zero-managed workspace history and explicitly does not replace Git or backups; host-side and external-system effects may not be reversible.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Agent Zero v2.14 Highlights - Significantly faster startup and chat loading — Embedding warm-up now overlaps with chat restoration and startup watchdogs, reducing median boot-to-first-token by ~27% and chat-switch latency by running model/profile refreshes independently of snapshot rendering. - Upgraded LiteLLM (1.104.0) and OpenAI SDK (2.54.0) with an automatic startup migration for existing Docker instances, including installed-SDK HTTP test coverage. …View on GitHub ↗ -
Release notes excerpt
Agent Zero v2.13 Highlights This release brings a significantly improved file browser, sidebar project folders, slash commands for messaging integrations, native commentary streaming, broad performance optimizations, and numerous quality-of-life and stability fixes across the WebUI, LLM transport, plugin system, and desktop environment. --- File Browser & Editor - Redesigned file browser navigation — configurable path bar with breadcrumb and text modes, folder autocomplete with keyboard navigation, and fully clickable file rows without a loading overlay …View on GitHub ↗ -
Release notes excerpt
Agent Zero Release Notes Connector & Transport Hardening - Enforced negotiated WebSocket payload ceilings with a symmetric checksummed transfer protocol for large operations, including bounded transfer state, disk spooling, and automatic cancellation on pause, reset, delete, or disconnect - Made workspace HTTP file transfers atomic with streaming uploads into partial files, size verification, fsync, and SHA-256 download receipts …View on GitHub ↗