OpenBot: OpenBot self-hosted AI agent platform
OpenBot is a self-hosted AI agent platform with a credible governance-first design, but it is still a beta template rather than a finished product. It is worth trying for controlled internal experiments where browser actions, approvals, and auditability matter more than a turnkey hosted service.
Why run AI agents on your infrastructure
Teams running AI agents near company tools need more than a chat interface: agents need isolated browser sessions, file workspaces, explicit permissions, intervention points, and a record of what they did. OpenBot addresses that operational gap with infrastructure you can host and inspect.
What OpenBot provides
OpenBot from CopilotKit is a self-hosted platform for AG-UI agents that act like private coworkers. Each coworker can receive a browser, files, and granted tools, while a single gateway evaluates actions and writes audit records. It supports your own model credentials, PostgreSQL, Docker Compose, React components, and MCP integrations.
Governed tools and isolated computers
- โEach coworker can run with its own browser profile, file workspace, and granted tools; isolated Bot computers can also run under gVisor where supported.
- โBrowser, file, shell, MCP, and component actions pass through one gateway that evaluates policy and records the action before execution.
- โCEL policies can inspect tool, actor, Bot, page, element, command, file, MCP, and initiator context; deny rules run first and broken policies fail closed.
- โThe UI can show a Bot's screen, activity, and saved-file metadata, while humans can take control when login or two-factor prompts appear.
- โAG-UI endpoints can be registered from `/bots`, with optional write-only authorization headers and exact host allowlisting for private endpoints.
- โMCP tools, deployment and personal skills, scheduled routines, encrypted credentials, and an administrator audit view are included in the platform surface.
- โAuthentication can use Google, Microsoft, Okta, SAML, or OpenID Connect, and administrators can control people, approvals, boundaries, plugins, and enterprise computer access.
Ways teams can use OpenBot
- โขGive an internal assistant a browser and private files for repetitive research or form-filling tasks while keeping actions inside a reviewed gateway.
- โขRun company-specific coworkers for expense-policy review, meeting follow-ups, release-note drafting, ticket triage, onboarding, interview write-ups, and renewal briefs.
- โขConnect an existing AG-UI agent from LangGraph, Mastra, CrewAI, Pydantic AI, Google ADK, or a custom implementation without making that framework the platform boundary.
- โขTest an agent that needs Google Drive, Notion, public-web research providers, or other MCP-backed tools with explicit grants and reviewable refusals.
- โขEvaluate an on-premises AI assistant with human takeover for login walls, two-factor prompts, and other actions that should not proceed unattended.
Install and configure OpenBot
OpenBot documents a Docker Compose-based setup. The prerequisites are Docker, Bun 1.4+, a CopilotKit Intelligence project and license, and a model key. Start with `cp .env.example .env`, run `bun install`, authenticate or select an Intelligence project with `npx --yes copilotkit@latest login` and `npx --yes copilotkit@latest project select`, run `bun scripts/setup-learning.ts`, fill `OPENAI_API_KEY`, and launch `bash scripts/start.sh`. The local app opens at `http://localhost:3010`. A published image is also available through `ghcr.io/copilotkit/openbot:latest`.
Strengths
- โGovernance is implemented in the execution path rather than left to agent instructions: actions are checked and recorded before reaching the computer.
- โThe platform is framework-neutral at the AG-UI boundary, so teams can retain their preferred agent implementation.
- โDocker Compose, a single image, PostgreSQL, and a documented start script make local evaluation reasonably concrete.
- โHuman takeover, deny-first CEL policy, encrypted credentials, and readable audit records address practical agent supervision concerns.
- โBrowser, file, shell, MCP, component, and routine surfaces are designed as one operating environment instead of disconnected demos.
Current beta limitations
- โณOpenBot is explicitly beta and under active development, so rough edges, bugs, and changing behavior should be expected.
- โณThe repository is a template: there is no hosted signup and no published package to depend on; teams must replace the example tenant package and operate the result themselves.
- โณA working setup needs Docker, Bun 1.4+, CopilotKit Intelligence and a license, model credentials, and an application key; the local Intelligence option is a macOS preview rather than a production deployment.
- โณThe repository states that every workspace is private, and the example configuration is intended for a starting point rather than an out-of-the-box operating product.
- โณPrivate agent endpoints require exact host entries, so onboarding internal agents can add configuration work and requires careful network review.
Hosted assistants and agent frameworks
OpenBot questions and answers
OpenBot is CopilotKit's self-hosted AI agent platform. It runs AG-UI agents with isolated browser sessions, files, shell access, MCP tools, approvals, and audit records on infrastructure you control.
OpenBot has no hosted version to sign up for. The repository is a template intended to be cloned, configured with your own coworkers and skills, and run on your own machine or infrastructure.
OpenBot can run any endpoint that speaks AG-UI, including agents written with LangGraph, Mastra, CrewAI, Pydantic AI, Google ADK, or custom code. Endpoints are validated for target and redirect rules.
OpenBot is designed for company-controlled infrastructure. Its Docker Compose setup and published image can run with embedded PostgreSQL, or you can point it at a database you already operate.
OpenBot is distributed under the MIT License, which permits commercial use, modification, distribution, and private deployment subject to the license terms. The repository is hosted on GitHub by CopilotKit.
OpenBot is currently beta and still under active development. The README explicitly warns about rough edges and bugs, so production readiness depends on your deployment, hardening, testing, and operational process.
Who should try it โ and who should skip
Try OpenBot if your team wants to experiment with self-hosted AI agents that can operate a browser, use MCP tools, and produce an auditable activity trail under your own deployment. Skip it if you need a polished hosted service, a stable release, or a dependency you can install and update through a package registry; this repository is a beta template that expects real implementation and operational ownership.
Related repositories
Curious whether OpenBot is right for you?
Let ChatGPT, Claude, or Perplexity look into it โ click below and see what AI actually says about OpenBot.
