What CopilotKit shipped
CopilotKit makes AG-UI, a protocol that connects agent back ends to user interfaces. On 9 November the company tagged OpenBot v1.0.0. OpenBot is a self-hostable workspace. You give a task to a "Bot", watch the Bot work inside its own container, and approve its next step in the same chat. The application code is MIT-licensed. MarkTechPost's write-up gives the same release date and licence. CopilotKit sponsored that write-up, so we've based the steps below on the repository and release notes.
Version 1.0 lets you connect more than one account of the same app through the Composio and Arcade connector services, and approval prompts now name the account involved. The release also adds "Teach a Bot", which turns a recorded task into a learned skill you can review and try. Bots now answer in Markdown by default, and the desktop app tells you which Docker resource is short (CPUs, memory or disk) and how to fix it, according to the release notes.
How the gateway works
For builders, the gateway is the part worth studying. The README says every browser and file action a Bot takes goes through a server-side gateway. The gateway resolves the target, evaluates policy and writes an audit row before it calls the Bot's computer, and the README states that no code path acts before that record exists. Policies are written in CEL (Common Expression Language, a small language built for evaluating rules). The rules fail closed: deny rules are checked before allow rules, a missing policy permits nothing, and a rule that errors refuses the action. Rules can read variables including tool.name, intent, bot.id, page.host, command, file. and mcp..
Each person gets their own container with a separate browser profile and workspace. The container binds to 127.0.0.1 and needs a per-container token. Setting COMPUTER_RUNTIME=runsc runs those containers under the gVisor sandbox runtime, and COMPUTER_SANDBOX=on turns on Chromium's own sandbox where the host allows it (README). The stack runs as separate services: the app on port 3010, the API server on 3001, the agent computer on 4100 and PostgreSQL on 5432. CopilotKit Intelligence, the company's back end, runs outside the stack.
Setting OpenBot up locally
According to the README, you'll need Docker, Bun 1.4 or later (some tests fail on Bun 1.3), a CopilotKit account (the free plan covers local use) and a model API key. The release notes set a minimum of 4 CPUs, 4 GB of memory and 40 GB of free disk for Docker. On Windows, the README says to run everything inside WSL 2 with Ubuntu.
- Clone the repository and install dependencies:
git clone https://github.com/CopilotKit/OpenBot.git
cd OpenBot
cp .env.example .env
bun install- Connect a local CopilotKit Intelligence instance:
npx --yes copilotkit@latest login
npx --yes copilotkit@latest local setup
npx --yes copilotkit@latest local connect --approve-connection- Edit
.env. SetOPENAI_API_KEY, or setANTHROPIC_API_KEYtogether withBOT_PROVIDER=anthropic. Then generate your ownKEY_ENCRYPTION_KEY:
openssl rand -base64 32- Start the stack with
bash scripts/start.sh, open http://localhost:3010, and stop the stack withbash scripts/stop.sh.
The README also documents a single-container option. It runs the published image with embedded PostgreSQL and maps port 3001.
Bringing your own agent and models
Any AG-UI endpoint can act as a Bot. You create the Bot from /bots with a name, instructions, an optional endpoint and an optional write-only auth header (README). OpenBot refuses agent endpoints on private addresses unless they appear in AGENT_ENDPOINT_ALLOWED_HOSTS. That variable takes a comma-separated list of exact hosts with optional ports. The configuration docs say the server won't start if an entry is a URL or contains a wildcard. CopilotKit's product page lists supported frameworks including the Claude Agent SDK, OpenAI Agents SDK, LangChain, Google ADK, Mastra and CrewAI.
The default model provider is OpenAI. Anthropic and Google are selected through BOT_PROVIDER. The variables OPENAI_BASE_URL, ANTHROPIC_BASE_URL and GOOGLE_GENERATIVE_AI_BASE_URL send model calls through a gateway or proxy of your choice (README). A team that already routes model traffic through an internal proxy for logging can keep doing so.
Writing a policy
You pass the action policy as JSON in AGENT_COMPUTER_POLICY. The configuration docs give this shape:
{"mode":"enforce","deny":[...],"allow":[...]}mode accepts enforce or dry-run, and invalid JSON or an invalid policy stops the server from starting. The configuration page doesn't define the fields inside a deny or allow entry, and it doesn't say what dry-run mode records. We couldn't find a worked rule example in either the README or the configuration page. Before you rely on dry-run output, test a policy against a throwaway Bot and read the audit log to see what was recorded.
Limits and costs
The MIT licence covers the OpenBot application only. The README says CopilotKit Intelligence has its own licence: a free local licence for personal use, a trial for Kubernetes evaluation, and a paid production licence for anything beyond those. So a production deployment depends on CopilotKit even though you can read and change the OpenBot code. The Learning feature needs an OpenAI or Anthropic API key, or a Claude or ChatGPT plan (release notes).
The release notes also flag some upgrade problems. When you connect a second account of an app, personal "allow without asking" rules stop applying and Bots start asking again for each account. Rolling back to 1.0.0-rc.1 or 0.3.4 after connecting multiple accounts isn't seamless, and the notes recommend restoring the pre-upgrade backup instead.
Pricing is free for personal use on a laptop, and you pay your model provider separately. OpenBot for Business costs up to US$29 per active user per month on pay as you go, and an active user is someone with at least one production interaction that month. Both paid tiers add a fee of 5% on monthly model spend above US$5,000. The pricing page's own example is US$10,000 of spend producing a US$250 fee. CopilotKit says the company doesn't resell inference or mark up tokens. All prices are in US dollars, so New Zealand teams will pay a varying amount in NZD.
What builders could do with OpenBot
The free local edition is the cheapest way to try an agent that can drive a browser under a written policy. Each action leaves an audit record that you can show a privacy officer or client. A useful first project is a support-ticket triage Bot that drafts a routing decision and waits for approval before posting. CopilotKit lists ticket triage as a use case in the MarkTechPost piece. For personal use, workspace data and API keys stay on your own machine, which makes it easier to keep client data off third-party hosting during a trial.
If you build a policy set or an AG-UI agent that works well with OpenBot, you can share it with other members on our projects page.
Sources
- Release v1.0.0, CopilotKit/OpenBot (GitHub)
- CopilotKit/OpenBot repository and README (GitHub)
- OpenBot configuration docs (GitHub)
- OpenBot pricing
- OpenBot product page
- CopilotKit Releases OpenBot 1.0 (MarkTechPost)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.