Agent Zero
Also known as: A0, Agent Zero AI, agent0ai
Open-source multi-agent framework that gives its agent a full Linux container with desktop and browser, subordinate agents, persistent memory, and scheduled tasks.
Agent Zero is an open source agentic framework that gives an AI agent a full Linux system inside a single Docker container, rather than a fixed set of prebuilt tools. The agent uses that environment as its workbench: writing and running code, driving a real XFCE desktop and a built in browser, editing documents and spreadsheets alongside the user, and creating its own tools on the fly. Almost nothing is hard coded. Prompts, tools and behavior all live in editable files, and any agent can spawn subordinate agents to break a task into focused subtasks. Memory persists across sessions, globally or per project, and scheduled tasks let the agent run recurring checks without being asked.
It works with OpenAI, Anthropic, Google and many other providers, plus local models via Ollama or LM Studio. It extends through MCP tools, skills and a Plugin Hub of more than 200 community plugins, and it runs self hosted so teams keep full control over data and credentials. Outside programs can also hand it tasks over an A2A endpoint. First released in 2024 by Jan Tomasek and developed by Agent Zero, s.r.o. of Brno, with releases every week or two and more than 19,000 GitHub stars, it targets developers, researchers and technical teams who want a transparent, fully customizable alternative to closed agent products.
Vendor details
Canonical URL
https://www.agent-zero.ai
Category
Multi-agent platform
Subcategory
Open-source agent framework
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Model-agnostic: works with OpenAI, Anthropic, and Google models, local models via Ollama, and account-backed plans through OpenAI Codex OAuth. Extends through MCP servers it connects to as a client and a Plugin Hub of 100+ community plugins, plus a built-in browser and SearXNG private search.
In practice
You hit a task no pre-built tool covers. Agent Zero writes and runs the code it needs inside its own Linux container, building the tool on the fly instead of waiting for a vendor to add a feature.
A job is too big for one agent to hold in its head. Agent Zero spawns subordinate agents, each taking a focused sub-task, and coordinates them back into one result.
You can't send company data to someone else's servers. Agent Zero runs self-hosted in a Docker container you control, so prompts, tools, and credentials stay on your own infrastructure.
Sources & related URLs
Related / legacy domains
Agentic Index coverage score
10.5 / 14 capabilities · 75%
| Integrations & Tool Calling | Full |
|---|---|
|
The agent connects to external tools through MCP servers by command or URL, installs capabilities from a Plugin Hub of 100+ community plugins with a security scan before install, writes and runs code to build tools it lacks, and reads credentials from named project secrets. Sourceraw.githubusercontent.com/agent0ai/agent-zero/main/docs/guides/mcp-setup.mdread 2026-09-27 |
|
| Workflow Orchestration | Full |
|
Every agent can create subordinate agents to break down work. The superior gives tasks and receives reports while each subordinate keeps its own focused context, v2.10 added durable subordinate lifecycles, and agent profiles set each agent's working style and tool policies. Sourcegithub.com/agent0ai/agent-zeroread 2026-09-27 |
|
| Knowledge Grounding & RAG | Partial |
|
The memory dashboard lists imported knowledge chunks beside remembered entries, and Agent Zero offers agentic RAG, with a self hosted SearXNG search for research. Agent Zero does not say how customer documents are ingested, indexed or refreshed. Projects store files the agent can work with rather than an index. Sourceraw.githubusercontent.com/agent0ai/agent-zero/main/docs/guides/memory.mdread 2026-09-27 |
|
| Human Oversight & Guardrails | Partial |
|
The user can pause and intervene in a running chat, Time Travel keeps snapshots of the workspace to inspect diffs and revert, agent profiles scope which tools and skills an agent may use (v2.9), and plugins are security scanned before install. There is no step where a person approves an action before it runs. Sourcegithub.com/agent0ai/agent-zeroread 2026-09-27 |
|
| Security, Identity & Governance | Not documented |
|
There is no SSO, identity integration or role model for who in an organization can use an instance, and no attestation or trust page. Named project secrets, plugin security scans and per profile tool scoping control what the agent touches, and the agent runs on the customer's own machine. Sourceraw.githubusercontent.com/agent0ai/agent-zero/main/docs/guides/usage.mdread 2026-09-27 |
|
| Observability & Auditability | Partial |
|
The UI streams each agent's reasoning and tool use, prompts and tools sit in inspectable files, Time Travel shows workspace diffs per snapshot, and a context window indicator arrived in v2.11. There is no run trace export, audit log or metrics view. Sourcegithub.com/agent0ai/agent-zeroread 2026-09-27 |
|
| Memory & State Persistence | Full |
|
Memory keeps facts, solutions and preferences across sessions so future chats do not start from zero. It is scoped as global or per project, so one client's memory does not leak into other work, and it is organized into main, fragments, solutions and skills areas, with a dashboard where the user can search, edit and delete single entries. Sourceraw.githubusercontent.com/agent0ai/agent-zero/main/docs/guides/memory.mdread 2026-09-27 |
|
| Deployment & Data Residency | Full |
|
The framework runs on the customer's own hardware in a Docker container, installed through a desktop launcher, a CLI or headless install, or on a remote server, and can use local models through Ollama or LM Studio so nothing leaves the machine. Sourceraw.githubusercontent.com/agent0ai/agent-zero/main/readme.mdread 2026-09-27 |
|
| Prebuilt Agents, Templates & Packs | Full |
|
The Plugin Hub on agent-zero.ai lists 215 community plugins that customers browse and install from inside the product, 66 of them tagged Agents, such as a BMAD set of 20 specialist agents. Skills and agent profiles also package reusable behavior. Sourceagent-zero.ai/p/pluginsread 2026-10-25 |
|
| Triggers & Channel Coverage | Full |
|
Tasks let the agent run work later, repeatedly or on demand, in scheduled, planned and ad hoc types, including recurring checks and monitoring tasks run with project scoped context and credentials, so the agent wakes on a schedule without a person starting it. Other work arrives through the web UI, desktop launcher and A0 CLI. Sourceraw.githubusercontent.com/agent0ai/agent-zero/main/docs/guides/usage.mdread 2026-09-27 |
|
| Model Flexibility & Routing | Full |
|
The user picks models from OpenAI, Anthropic, Google Gemini, Azure, xAI, Mistral, DeepSeek, Groq, OpenRouter, GitHub Copilot, Hugging Face and local Ollama or LM Studio, sets model presets per project, and can sign in with an OpenAI Codex account. Sourceagent-zero.airead 2026-09-27 |
|
| APIs, SDKs & MCP Extensibility | Partial |
|
Plugins, Skills, custom tools and code level extensions change the framework, and the A0 CLI drives it from the host. Agent Zero also runs as an A2A server: an outside program can send it a task at http://HOST:PORT/a2a/t-API_TOKEN, optionally scoped to a project, with the token generated from the user's credentials. No agent card is published for that endpoint. Sourceagent-zero.ai/p/docs/mcp-a2aread 2026-10-25 |
|
| Testing, Debugging & Optimization | Partial |
|
Debugging aids ship, the visible reasoning stream, Time Travel diffs and revert, and a Context Doctor and context usage indicator in v2.11. There is no harness for scoring or comparing agent versions. When the agent tests its own code, that is part of the user's task rather than a check on the agent. Sourcegithub.com/agent0ai/agent-zero/releasesread 2026-09-27 |
|
| Browser & Computer Use | Full |
|
The agent drives a full Linux XFCE desktop and its GUI applications and a native browser with DOM annotation for clicking, typing and capturing elements, rebuilt in v2.10 as an interactive shared viewport, alongside shell and LibreOffice document cowork. Sourceraw.githubusercontent.com/agent0ai/agent-zero/main/readme.mdread 2026-09-27 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Per action tool permissions in Agent Zero v2.15 let an agent profile allow or deny single operations inside the 12 bundled tools that have several actions, so a file tool can read while writing stays blocked. The Agent Editor shows these as grouped action controls in both Easy and Advanced modes. Settings changes over Telegram and WhatsApp now also require named administrators, so an accepted chat contact can no longer lift their own restrictions.
Bears on: Human approval / guardrails
View sourceRelease 2.14 of Agent Zero runs direct delegation to a subordinate agent through the child's own context task, so child chats show as Working or Paused and can be stopped on their own. Cancelling the caller now carries through to the child, while stopping only the child can be recovered, and the sidebar shows subagents at every delegation depth.
Bears on: Human approval / guardrails
View sourceAgent Zero v2.13 replaces runtime=input with input_remote for remote execution, without a compatibility alias. It also adds Telegram and WhatsApp slash commands with project aware command catalogs and scoped integration settings, plus a configurable browser JavaScript evaluation timeout.
Bears on: MCP / tool calling / API
View sourcePricing
Free and open source
No billing unit. The framework is free, and the user pays their own model provider and compute.
What is public
Agent Zero is a free open source framework run on the user's own machine or server. No paid plan is published. The site also links an LLM API (beta) page, and its terms and price are not stated.
Cost watchouts
Model provider usage is the running cost and scales with agent activity, especially with subordinate agents and long tasks. Local models remove it at the cost of your own hardware.
Variable cost rationale
Agent Zero meters nothing, since the framework is free and self hosted. What still costs money is model provider usage and compute, which the user pays directly and which grow with agent activity, subordinate agents and long tasks. Local models through Ollama or LM Studio remove the model bill in exchange for running on the user's own hardware. The framework itself adds no spending cap.
Sales call required
No, self serve available
Free / trial
Free and open source
Lowest paid plan
No paid plan published
Missing data
The terms and price of the LLM API (beta) offering.
Related vendors
- CrewAI — Open source multi agent framework (Python, MIT) of role based Crews…
- Relevance AI — Platform for building and deploying a workforce of AI agents with a…
- Akka — Enterprise agentic AI platform that builds and runs multi agent…
- Artian — Enterprise multi agent AI platform for financial services, where a…
- Databricks Agent Bricks — Databricks' agent platform, now sold as Agent Bricks, for building,…
- Dataiku — Enterprise AI and machine learning platform with an agentic layer…
Alternatives to Agent Zero
The closest documented capability profiles to Agent Zero among multi-agent platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Akka12.5 / 14Adds documented Security, Identity & Governance
- Helio8.5 / 14Adds documented Security, Identity & Governance
- Pokee AI8.5 / 14Adds documented Security, Identity & Governance
- SimplAI12.5 / 14Adds documented Security, Identity & Governance
- Artian11.0 / 14Adds documented Security, Identity & Governance
- CrewAI13.0 / 14Adds documented Security, Identity & GovernanceAgent Zero vs CrewAI →
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded