CopilotKit's OpenBot gives each agent its own browser and logins, with every action decided first and logged after
Synopsis
CopilotKit released OpenBot, a cloneable template for an agent workspace that runs on your own infrastructure. It admits any agent speaking the AG-UI protocol as a coworker with its own browser, its own logins and its own files, and every action an agent takes on a computer, a file, an MCP server or a component passes through one gateway that decides before and records after. A deployment starts with two coworkers defined as configuration rather than code, with ten more example roles waiting to be copied in; the project is in beta, has no hosted version and is not published as a package to depend on.
Interpretation
A company can take an agent workspace that runs on its own machines instead of subscribing to one someone else operates. Comparable assistants are hosted by the provider and cannot be changed inside; OpenBot ships as a template, every workspace in the repository is private, there is no hosted version to sign up for and nothing is published as a package to depend on. The stated path is to clone the repository, replace the example tenant package under examples/ with your own coworkers, channels and skills, and run it, which the project describes as a starting point rather than a finished thing somebody operates for you.
Agents are managed as coworkers with their own working environment rather than as code that can call tools. Each coworker gets a real browser, its own logins and its own files, and only the tools you grant; a person can watch it work on its own screen, take the wheel when it reaches something it should not do alone, then hand it back. Anything a Bot does to a computer, a file, an MCP server or a component goes through one gateway that decides and records it.
Admission is not tied to a framework, and governance rides the protocol rather than the framework. A Bot is any endpoint speaking AG-UI, the open protocol for agent-to-user interaction; agents built with LangGraph, Mastra, CrewAI, Pydantic AI, Google ADK or written by hand all arrive the same way. A deployment starts with General Assistant and Knowledge, which are configuration rather than code; a Risk Analyst joins once its endpoint is configured, and ten more roles sit in examples/fintech/catalog/, each doing one job.
Perspective
The material describes a laptop-scale starting point: Docker Compose brings up every part, data sits in your own PostgreSQL, an administrator supplies the model credential, and a fresh clone admits every request as one administrator until sign-in is turned on for others to reach the deployment.
The project describes itself as early, in beta and under active development, with rough edges and bugs expected and things expected to move; the material reports no performance or security evaluation.