CopilotKit Releases OpenBot 1.0: An MIT-Licensed, Self-Hostable AI Coworker With Built-In Governance

CopilotKit, the team behind the AG-UI protocol, released OpenBot 1.0 on November 9. It is an MIT-licensed agent workspace that companies can run on their own laptops, cloud accounts or Kubernetes clusters, using their own models and agent frameworks. Each bot gets its own container with a browser, logins and files, and reaches only the tools it is granted.

One disclosure up front: MarkTechPost notes the article is sponsored by CopilotKit. We are treating the claims as the vendor's description of its product, not independent testing.

We think the governance design is the interesting part. Every browser click, file write, MCP tool call and rendered component passes through a single gateway that evaluates a policy, writes the audit record first, and only then acts. Deny rules run before allow rules, and a broken rule refuses rather than opens. That fail-closed posture is exactly what agents touching real systems should have. Too many agent products bolt on oversight after the fact. Putting it in the action path is the better pattern, and we'd like to see it become standard.

The bring-your-own approach also matters. OpenBot works with any agent that speaks AG-UI, with listed support for frameworks including the Claude Agent SDK, OpenAI Agents SDK, LangChain, Google ADK, CrewAI and others, and with models from OpenAI, Anthropic, Google and xAI as well as open models like DeepSeek, Qwen and Mistral. For companies wary of handing data and rules to one vendor, that flexibility is a real advantage over closed hosted coworkers.

Now the risks. An agent with its own browser and saved logins is a powerful thing to deploy. Policies are only as good as the people writing them, and a misconfigured grant can quietly give a bot reach into systems nobody intended. Version 1.0 also adds a feature that records a task once and turns it into a reusable skill, and lets bots hand work to other bots. Those are useful, and they also multiply the paths where a mistake can travel. The business tier adds human approval routing, spending limits and audit export, which helps, but those controls live in the paid plan.

The pricing model is worth knowing: free for personal use, up to $29 per active user per month for business, and a 5% charge on model spend above $5,000 a month. Open code with paid governance is a familiar playbook. It can work well, as long as the free version doesn't train people into deploying agents without guardrails.

Our take: this points in the right direction for workplace AI, where agents do real work under clear rules and leave a trail. We'd want independent security reviews before trusting it with anything sensitive.

Example uses offered include support ticket triage with human approval before posting to Slack, expense claim review against policy, and drafting operational fixes for human sign-off. The release adds desktop apps for Mac and Windows, multiple accounts per app through Composio and Arcade, and early access for iOS, Android, Slack and Microsoft Teams. Companies deploy via a single Helm chart on AWS, Azure, Google Cloud or any Kubernetes cluster.