OpenBot: Govern and Run AI Coworkers on Your Infrastructure

Overview

OpenBot is a TypeScript application for deploying AI coworkers inside your own infrastructure. It connects agents that speak AG-UI, rather than requiring a particular agent framework, and gives each one a separate computer environment for browser and file work.

Its central purpose is to make agent actions governable: browser, file, shell and MCP tool calls pass through a gateway that applies policy and records an audit entry before acting. It is aimed at teams that want to adapt and operate their own agent platform, not users seeking a hosted service.

Key Features

• Connect custom AG-UI agents, including agents built with different frameworks or by hand.
• Give each coworker an isolated computer with its own browser profile and workspace.
• Apply CEL-based policies to computer actions, with deny rules evaluated before allow rules.
• Record permitted, refused and failed actions in an audit interface.
• Configure MCP servers, credentials, components and skills with access controls.
• Support scheduled routines, persistent conversations and memory through CopilotKit Intelligence.
• Offer a web interface for managing coworkers, channels, approvals and administration.

Use Cases

• Platform or IT teams can deploy an agent workspace within company infrastructure and control access to tools and data.
• Security teams can define which browser, file or shell actions agents may perform and review the resulting audit trail.
• Developers can connect an existing AG-UI agent to a shared interface without adopting one specific agent framework.
• Teams can create purpose-specific coworkers for tasks such as ticket triage, expense review or meeting follow-ups, then adapt the example configuration.

Getting Started

The README describes prerequisites and environment configuration. For a local setup, start with:

cp .env.example .env
bun install

Then configure the required Intelligence and model credentials and run bash scripts/start.sh. See the README for the complete setup and deployment instructions.

Alternatives

• BrowserOS: BrowserOS focuses on AI-driven tasks in a Chromium browser, while OpenBot provides self-hosted coworkers with broader tools, files, and configurable action policies.
• lobehub: LobeHub focuses on building and coordinating agent workflows, while OpenBot gives agents their own browser and files with policy controls for computer actions.
• astron-rpa: AstronRPA builds scheduled desktop and web automations with a visual workflow editor, while OpenBot centers on AI coworkers operating through tools and a browser.

Considerations

• The project identifies itself as alpha and under active development, so expect rough edges and changes.
• OpenBot is a template to clone and customize, not a hosted product or a package intended to be installed as a dependency.
• A local setup requires Docker, Bun 1.3 or later, a CopilotKit Intelligence project and a model key. The example proof-of-concept Bot uses OpenAI.
• The README describes a single-user administrator mode for fresh local setups. Configure an identity provider before allowing other users into a deployment.
• Browser computers handle real logins. Keep their services private and configure secrets and action policies carefully.