CopilotKit OpenBot is a workspace for AI coworkers that runs on infrastructure you own. Since version 1.0.0 shipped on November 9, 2026, it comes two ways: a desktop app for Mac (Apple silicon and Intel) and Windows that you install without a terminal, and the MIT licensed template you clone, customize, and connect to your own AG-UI agents. For teams, CopilotKit now offers OpenBot for Business in early access, with company sign in, approvals, audit, and help with hosting.
That framing matters. Earlier writeups, including the first version of this one, described OpenBot as an Alpha template with no signup page. That is no longer the picture. What hasn't changed is the core: a customizable agent app, a computer for the Bots to work in, governance hooks, and example coworkers you adapt. Below is what it is, what you need to run it, and when it is (and is not) the right call.
What OpenBot is (and is not)
OpenBot gives your company a ChatGPT-like surface that runs on your stack. Employees chat with bots. Those bots can use tools, browse, and work through multi-step jobs. The difference that matters: it is designed so you own the runtime, the logs, and the policy layer.
It is:
- A desktop app plus an open template. The personal edition installs on your Mac or Windows laptop. The MIT licensed template is what you fork, white label, or customize for your company.
- Built around AG-UI agents. You bring agents from stacks like LangGraph, Mastra, CrewAI, Pydantic AI, Google ADK, or hand-written agents that speak AG-UI.
- Oriented to AI coworkers: bots that act inside company systems with clearer boundaries than a naked chat window.
It is not:
- A finished product you turn on for the whole company with one click. Web, iOS, Android, Slack, and Teams frontends and OpenBot for Business are still in early access.
- A promise that "free" means zero cost. You still need a model key, and anything past a laptop or an evaluation needs a production license (details below).
AG-UI first: bring your own agent
OpenBot's center of gravity is the agent you connect, not a single locked-in model brand.
AG-UI (Agent-User Interaction) is the protocol lane the project centers on. In plain terms: your agent talks to the OpenBot app through a shared interface, so the chat UI, approvals, and tool gateway can stay stable while you swap agent frameworks.
That is why official docs list several agent options instead of one proprietary brain. If you already have an agent harness at work, OpenBot is closer to "put a company-facing app around it" than "throw away your stack and start over."
Computer per agent, not one shared browser tab
The README describes a computer per person: each person gets their own container, with a separate folder and browser profile for each of their Bots. Mark a Bot as Isolated and it gets a container of its own. That is stronger than "a sandboxed browser somewhere on a shared host."
Why it matters for real work:
- One Bot's messy login state should not bleed into another Bot's browser profile.
- You can reason about permissions per coworker, not as one giant shared desktop.
- A fail-closed gateway sits in front of tools, so actions that are not allowed should stop before they run.
OpenBot also leans on CopilotKit Intelligence for durable conversations and memory. Intelligence is licensed separately from the MIT template, so treat it as part of the runtime story, not an optional footnote.
What you actually need to run it
It depends on which path you take. From the official README (confirm on the current version before you start):
On your laptop (desktop app). Download the notarized .dmg for Mac or the signed Windows installer. The README says no terminal or developer tools are needed: the app combines the OpenBot frontend with a Docker based backend, and setup handles the local Intelligence license for you. You still bring a model key.
From source (the template).
- Bun 1.4+ (the README notes some tests fail on 1.3)
- Docker
- A CopilotKit Intelligence project and license. A free plan is available, and Intelligence can be self hosted.
- A model key. No model ships in the box. OpenBot encrypts stored model credentials and keeps them out of logs.
A fresh clone starts in single user mode: .env.example carries OPENBOT_SINGLE_USER=true, which admits every request as one administrator so you can try it on a laptop. Sign in is required before anybody else can reach the deployment. On Windows, the README asks you to run the source version inside WSL 2 with Ubuntu.
For a team (Kubernetes). The README now leads its deployment section with Kubernetes: AWS EKS, Azure AKS, Google GKE, or your own cluster. The Helm chart installs Intelligence alongside OpenBot by default, and a first evaluation needs a cluster with amd64 nodes, a default StorageClass, kubectl, Helm, a trial or production license, and a model key. A single Docker image is also published for a setup without Kubernetes.
Which Intelligence license you need, per the README:
| Where you run it | Intelligence license |
|---|---|
| Your laptop, personal use and experimentation | Free local license, handled by desktop setup |
| Kubernetes evaluation | Free trial license |
| Kubernetes in production | Production license (included with OpenBot for Business) |
If someone tells you it is "just MIT and you're done," that undersells the Intelligence license, the model key, and the ops work. The template may be open. Running it still has moving parts.
Two starter Bots and ten examples (fintech set)
The team template comes with two Bots, configured rather than coded: a General Assistant for everyday work and Knowledge, which answers questions from the company documents you connect and cites its sources. Ten more examples wait as YAML files in examples/fintech/catalog/, each doing one job, such as reviewing an expense claim, drafting release notes, triaging a ticket, or helping a new starter onboard. You copy one into examples/fintech/agents/ and restart OpenBot to try it, and you grant tool access separately, so a Bot can read a document without permission to edit it.
Treat them as starting points you customize, not as certified bots for your compliance team. They're useful for learning the shape and the prompts. They aren't a substitute for your own roles, tools, and approval rules.
What the README says it can do
Beyond chat and a computer per bot, the README lists a few features worth knowing about before you decide:
- Reach. A Bot can continue a conversation in Slack, Microsoft Teams, or by text message, not only in the web app.
- Routines. Ask a Bot to do something on a schedule and it runs as you, in the channel you asked in. There's a 15 minute floor and a cap of 20 enabled routines, and ten failures in a row switch a routine off. Routines need a worker process, which the start script runs locally.
- Governed tool access. The MCP catalogue ships Google Drive, Notion and public web research providers, and Composio brokers a few hundred more apps, each reached as the person asking. Rules written in CEL (a small policy language) can allow or deny actions, deny wins over allow, and a missing policy permits nothing.
- Sign in. Google, Microsoft or Okta, or your company's own SAML or OpenID Connect provider. Admins can list, promote and remove people, and an audit page shows what was permitted, refused and failed.
For the control side of the story, read our guide on OpenBot governance for enterprise.
A realistic setup path
- Decide your path: the desktop app to try it on your own laptop, or the source template if you plan to customize it.
- Install the desktop app, or clone the repo and follow the current quick start.
- Connect an AG-UI agent you already understand (or one of the documented options).
- Confirm your Intelligence project and license setup and your model key. Remember a fresh clone runs as one administrator until you set up sign-in.
- Create one bot with a tight tool allowlist and watch the computer boundary (container + browser) before you add more bots.
- When you outgrow the laptop, look at the Kubernetes path (or OpenBot for Business) and your company's usual production checklist (secrets, networking, audit retention, who can approve risky tools).
Skip the fantasy where you paste one command and wake up with a governed workforce of agents. Even at 1.0, that's not how it works.
When OpenBot is a good fit
Choose it when:
- You want coworkers on infrastructure you control.
- You already care about agent frameworks and want a company UI around them.
- You need per-bot computer isolation and a place to hang governance, not only chat.
Look elsewhere (or wait) when:
- You need a managed SaaS with support SLAs tomorrow.
- Your team cannot own Docker, secrets, and model spend.
- You expected a finished product catalog instead of two starter Bots plus examples.
How this compares to "just use a closed chat app"
Closed chat apps are fast to try. They are often weak when you need your data residency story, your tool gateway, and your own agent harness under one roof.
OpenBot flips that tradeoff: more ownership, more setup, and a 1.0 that is days old. Grok Bot and similar computer use agents are useful comparisons when you care about desktop control and agent autonomy, but they are different products with different hosting stories. Do not treat any of them as drop in clones of each other.
If you are following AI coworker and agent tooling more broadly, our AI tools coverage rounds up more breakdowns like this one, from AG-UI agent frameworks to browser and computer-use agents.
Want a structured way to practice?
If you are exploring AI coworkers and self-paced learning paths (agents, automation, content) without a fixed cohort, look at how 4Geeks AI Flex is structured: ready-made paths, an AI tutor, and mentorship you can start when you are ready.
Sources to recheck before you ship a pilot
- GitHub: CopilotKit/openbot (desktop downloads, requirements, licensing, deployment, example pack)
- GitHub releases: CopilotKit/OpenBot releases (v1.0.0, November 9, 2026)
- Plans: open.bot/pricing
- Product page: copilotkit.ai/openbot (positioning)
This page was last checked against the README and releases on November 11, 2026.
Release candidates for the next versions were already landing the day after 1.0, so README details still move fast. Read those pages again the day you install.
If you are looking for structured AI training, review our AI programs side by side: length, level and outcomes.