CopilotKit has released OpenBot 1.0, a free and open-source workspace that lets businesses run their own AI agents on their own infrastructure. Its headline trick is simple: show the software how to do a job once, and it learns to do it again. The 9 November release turns the familiar chatbot into something closer to a junior colleague.
Companies have spent two years experimenting with AI assistants that take actions, yet many hesitate to hand real work to systems they cannot inspect or restrain. OpenBot answers that hesitation directly. Released under the MIT licence, OpenBot runs on a laptop, a private cloud or a Kubernetes cluster, and lets a team choose its own models, agents and tools instead of accepting a single vendor’s offering.
How OpenBot Works
Working with OpenBot begins in one conversation: a user assigns a task, watches the Bot carry it out, reviews what it found and approves the next step. Each Bot works inside its own Chromium browser profile and workspace folder, and administrators can place individual Bots in isolated containers. A Bot can only reach the tools it has permission to use, so access is granted deliberately rather than by default.
Version 1.0 answers in Markdown by default and can present its work as charts, checklists, tables or cards. A support queue is far more useful as a structured summary than as a wall of prose.
Policy and Approval Controls
The feature that sets OpenBot apart is its policy gateway, which examines an action before it runs. Rules in the Common Expression Language vet browser clicks, file writes, Model Context Protocol calls and interface actions. Deny rules are checked first, every decision is recorded before an approved action runs, and an action is blocked when no policy covers it or a rule fails.
In practice, a Bot can sort support tickets and wait for human approval before posting to Slack, flag expenses that break policy, or draft a fix for review. The business edition adds single sign-on, SCIM synchronisation and an administration console, with Google, Microsoft Entra and Okta supported. Access is granted by Bot, role or person, and Bots act using tokens carrying only a subset of a user’s permissions. Spending reports, token limits and audit exports complete the package, and CopilotKit says the product is SOC 2 Type II compliant.
What’s New in OpenBot 1.0
The 1.0 release also polishes the everyday experience. The desktop app supports Windows and both Apple Silicon and Intel Macs, with signed updates and pre-upgrade backups. The Teach a Bot feature converts a recorded task into a reusable skill that a user can review before it is saved, turning one demonstration into a repeatable routine. Composio and Arcade let users link several accounts for the same application, each with its own approval rules, and Composio alone connects to several hundred apps. Bots can hand work to specialist Bots within administrator-set limits, with iOS, Android, Slack and Microsoft Teams in early access.
Supported Models and Agents
OpenBot connects to agents that support AG-UI, a protocol CopilotKit helped develop. Supported frameworks include the Claude and OpenAI agent SDKs, LangChain, Google ADK, Mastra, AWS Strands, the Microsoft Agent Framework and CrewAI. Teams can connect models from OpenAI, Anthropic, Google and xAI, and open models such as Kimi, DeepSeek, Qwen, Nemotron and Mistral are supported too. The sandbox, MCP hub and browser can each be swapped out, so no team is locked into one supplier.
OpenBot 1.0 vs. OpenAI Dots and OpenClaw
OpenBot does not arrive in an empty market. Against OpenAI’s Dots, the contrast is ownership: OpenBot is MIT-licensed and runs on a laptop, a private cloud or Kubernetes with a choice of models, while Dots is proprietary, runs on an OpenAI-hosted computer and uses GPT-6 Astra and GPT-6.1 Terra. OpenClaw is also MIT-licensed and free, but relies on its own built-in agent rather than OpenBot’s Bots and policy checks. OpenBot connects through Composio, Arcade and MCP, Dots claims more than 4,000 apps, and OpenClaw works through plugins and skills. OpenBot is free for personal projects; business plans cost up to 29 dollars per active user per month.
Deployment and Pricing
For personal use, OpenBot keeps workspace data and API keys on the user’s machine, and users pay their model providers separately. Companies can deploy it with a Helm chart on AWS, Microsoft Azure, Google Cloud or another Kubernetes cluster. Business pricing is unusual: CopilotKit charges 5 percent on monthly model spending above 5,000 dollars per organisation, while the annual plan adds air-gapped deployment and round-the-clock support.
That model tells its own story about who OpenBot is for: not casual users, but organisations that already spend serious money on AI models and want agents they can govern, audit and keep inside their own walls. Whether the bet pays off will depend on how quickly teams find teaching a Bot once genuinely easier than doing the job themselves. The age of the AI coworker has moved past the demonstration stage, and OpenBot 1.0 is one of the most complete attempts yet to make it safe enough for real work.
