What Vibehub is
Vibehub is an IDE for AI agents. Instead of one chat window, it runs many agents at once, each in its own session, each working on a real task against a real codebase. Sessions can spawn other sessions to delegate work, message each other, and report back. A Kanban board tracks the work, and starting a task hands it to a fresh agent that does the job and sets the card to review when it is done.
We built it for ourselves because we needed it. This site was built inside Vibehub, by agents running in it, coordinated from it.
The problem it solves
A single AI agent in a single window is a demo. The moment you want real output, the same failures show up every time. The agent loops without a stopping condition. It has no way to act on your systems, so it can only talk. It runs on one machine with no record of what it did, so you cannot audit or improve it. And it has no way to hand part of the job to another agent when the task is too big for one context.
Vibehub is the answer we built to those failures. Every one of them is a design decision in the platform, not a patch.
The architecture decisions that matter
Vibehub is TypeScript end to end: an Express API and a Node service on the back end, a React front end, and SQLite for local state so an instance runs without external dependencies. The code is split into modules so a new capability, an email integration, a scheduler, a browser driver, ships as a self-contained package rather than a change threaded through the whole app.
A few decisions are the ones that make it useful for real work:
Sessions are the unit, not chats. Each agent run is isolated and addressable. One agent can create another and wait for its reply, which is how a large task gets split across focused workers instead of overflowing one context window.
Tools go through the Model Context Protocol. Agents reach the outside world, version control, email, storage, a real browser, cloud APIs, through MCP servers rather than hard-coded glue. That is the same MCP development pattern we bring to client work, and running our own daily is why we trust it.
Humans stay in the loop by default. Any action that is hard to reverse pauses for approval. The agent proposes, a person confirms. That single rule is the difference between an assistant you can leave running and one you have to watch.
Everything is observable. Sessions, tasks, and activity are all recorded, so when an agent makes a call you disagree with, you have the trace to see why and fix it.
Why this is proof, not a pitch
Most studios that offer agent work are building their first one on your budget. We are not. Vibehub is a production system we depend on, and the patterns we use in it, scoped autonomy, MCP tools, approval gates, delegation between agents, are the exact ones we apply when we build AI agents for you. The platform is the portfolio.
If you have a job you want an agent to own, the people who designed and run Vibehub are the ones you will work with.