Your tasks. AI agents.
You approve the code.

Sign up, choose VB Box Pro and bring your Git repository. We provision and operate the development server, so you do not need to supply or maintain one. The agents use your project's own build, test and deploy commands.

No. There's a desktop app — Windows today — that opens the same board and finds the agent CLIs already installed on that machine — Claude Code, OpenCode, Codex, Gemini, Ollama and LM Studio, including ones installed inside WSL on Windows. A person can open a card and run it right there, under their own agent sign-in, and the card still walks the same columns and stops at the same human merge gate. Nobody shares an API key, and work that isn't allowed to leave the building can run against local models instead.

You choose the agents and models used for your project. With VB Box Pro, they work in your project checkout on the development server we manage for you. VB Box Enterprise is the option for deployment on your own infrastructure.

VB Box Pro runs on a development server we provision and operate for you. You get the board without needing to operate the server. If you require installation on your own infrastructure, contact us about VB Box Enterprise. Your selected AI provider processes model requests under its own terms.

Four agent CLIs today — Claude Code, OpenCode, Codex and Gemini — each with resumable sessions, offered from a per-board catalog an admin controls. Bring credentials for any of them, or use ours.

Yes. An agent can draw wireframes, architecture and decision diagrams on a real vector canvas through the board's own tools, then submit the design for review. Only a human can approve it — the agent has no tool that can approve its own design — and the build sub-task is only created once that happens.

No. Planning approval, an AI review phase and an AI test gate all happen first, and even then no MCP tool exists that lets an agent merge — that action is reserved for a human, every time.

A configured per-task and per-day USD budget is checked before a turn is even allowed to start; a run that would exceed it is stopped and sent back with an alert instead of quietly overspending. Cost falls over time because agent spend tracks loop length — how many iterations a turn takes and how much context gets re-read each time — and project memory, per-task memory and a code index all attack that directly, so the same class of task gets cheaper as the board's history grows. It's a mechanism you can check on your own board's economics panel, not a promise.

Yes — configurable parallel queue slots (1 to 20, serial by default). Turn on git-worktree isolation and every task gets its own working directory and branch, so concurrent agents never trample each other's checkout.

Yes. Point a board at an empty repository and the first cards are the scaffold — stack and database chosen deliberately on a card, not guessed from a prompt. We've run this end to end on one of our own apps: 50 tasks, 7 root tasks, all done, roughly ten days from the first card to a built app, its backend, and a marketing site.

They never touch VB Box directly — they work in the team's chats, planning and the Space wiki/board inside AiMixup. A tagged chat message, a planned card, or a wiki suggestion all land on the team's own Space Kanban; one explicit delegate action then hands a card to VB Box to build. Status mirrors back to the card the team is already watching, and finished work returns to the folder wiki as a Development archive. We wire the two surfaces together as part of setting up your instance — it isn't a self-service toggle you switch on yourself.

Jira and Linear track work a human does. VB Box's card is the work — a real agent session with a shell, a checkout and a turn budget behind it, not a field a human fills in. Full comparison: /compare/vb-box-vs-jira.

Sign up for an AiMixup account, choose VB Box Pro and start building. We provision and operate the development server. Contact our team only if you need VB Box Enterprise on your own infrastructure.