AgentOS

The operating layer for digital workers

Running one AI agent is easy.

Running many agents across real projects, accounts, models, schedules, files, and approval boundaries is an operating problem.

AgentOS gives a human operator one place to:

• create and organize digital workers;
• give each worker identity, context, memory, policies, tools, and model access;
• assign immediate, scheduled, recurring, or one-off work;
• inspect live activity, transcripts, outputs, token usage, and created files;
• connect workspaces to channels and browser-based accounts;
• review system health, capabilities, fallbacks, and failures;
• keep humans in control of sensitive actions.

The goal is not another chat interface. It is a practical operating system for AI-native work.

Why AgentOS

AI agents are becoming cheaper, more capable, and more persistent. The next bottleneck is no longer generating a response. It is designing and operating the organization around them.

Digital workers need more than prompts:

• Structure — clear workspaces, roles, goals, and ownership.
• Context — project files, memory, policies, accounts, and operating instructions.
• Execution — models, tools, channels, automations, and real runtime access.
• Coordination — tasks, handoffs, schedules, shared visibility, and recoverable state.
• Control — approvals, diagnostics, auditability, and human intervention.

AgentOS is being built for that layer: the place where a person can turn individual agents into a functioning digital team.

Product tour

Get started in 5 minutes

Deploy on Railway

AgentOS includes a production-oriented Railway deployment path that runs AgentOS and the pinned OpenClaw Gateway in the public application service and runs interactive Chromium in a separate private browser-worker service. The deployment provisions one persistent /data volume for AgentOS/OpenClaw state and a second dedicated /data volume for Chromium profiles, keeps the Gateway on container loopback, initializes Instance Protection from the username and password supplied during deployment, and waits for the private worker and Gateway before passing Railway's healthcheck. The Railway external supervisor owns the Gateway process lifecycle; AgentOS configures and observes OpenClaw through its native Gateway and can request a supervised Gateway restart through the lifecycle API without opening a host terminal.

The official published Railway template provisions both required volumes, generates the internal AgentOS, OpenClaw, and browser-worker secrets automatically, and wires the private worker URL through Railway private networking. The only required operator input is the initial administrator password. See docs/deploy-on-railway.md for the complete runtime, private-networking, and template update model.

After the first deployment:

1. Open the generated Railway domain and sign in with the initial admin credentials.
2. Connect a model/provider from AgentOS Setup Center.
3. Create the first workspace. AgentOS and OpenClaw state will persist across redeploys.

The initial password only creates the account when no Instance Protection state exists. Changing that Railway variable later does not reset or replace the administrator account; use AgentOS account settings to change credentials. After the first successful sign-in, you may remove AGENTOS_INITIAL_ADMIN_PASSWORD from Railway because the protected account is already stored on the volume.

Install

macOS or Linux:

curl -fsSL https://raw.githubusercontent.com/SapienXai/AgentOS/main/install.sh | bash

Windows PowerShell:

iwr https://raw.githubusercontent.com/SapienXai/AgentOS/main/install.ps1 | iex

Or install with a package manager:

pnpm add -g @sapienx/agentos
# or
npm install -g @sapienx/agentos

Install a specific published version:

macOS or Linux:

curl -fsSL https://raw.githubusercontent.com/SapienXai/AgentOS/main/install.sh | AGENTOS_VERSION=0.7.8 bash

Windows PowerShell:

$env:AGENTOS_VERSION='0.7.8'; iwr https://raw.githubusercontent.com/SapienXai/AgentOS/main/install.ps1 | iex

Start AgentOS

agentos start --open

Verify the local environment:

agentos status
agentos doctor --deep

Then:

1. complete OpenClaw and model onboarding;
2. create or select a workspace;
3. create a digital worker;
4. assign a mission or scheduled task;
5. inspect the runtime, transcript, and output.

If OpenClaw is already installed, AgentOS connects to the live Gateway and projects its real state. If OpenClaw is missing or unhealthy, AgentOS shows an explicit setup or unavailable state instead of presenting fake operational data.

What AgentOS does

Build digital workers

• Create and edit workers with role presets such as worker, setup, browser, monitoring, and custom.
• Configure identity, model, heartbeat, file access, install scope, network policy, and operating instructions.
• Maintain OpenClaw-native workspace context through AGENTS.md, SOUL.md, IDENTITY.md, and optional USER.md; keep command guidance in AGENTS.md under ## Tools, curated memory in MEMORY.md/memory/, and AgentOS project docs in docs/, deliverables/, and skills/. Heartbeat cadence is