VibeLoop
An orchestration platform for coding agents — Kanban board, live terminals, MCP server per agent, and real-time notifications so you know the moment an agent gets stuck and can act immediately. Agents run on AWS VMs sandboxed with ptracer and minivisor for security and full observability.
VibeLoop is a platform for running and supervising coding agents. The core problem it solves: agents get stuck, go silent, or take a wrong turn — and you don't find out until minutes later when you glance back at the terminal. VibeLoop surfaces that signal immediately so you can intervene, redirect, or unblock the agent before it wastes time.
How it works
Each agent session gets:
- A live terminal in the browser (xterm.js + node-pty) with real-time streaming over SSE
- A dedicated MCP server over stdio, exposing tools the agent can call:
createTodos,listTodos,createSummary,askQuestion, andaskQuestionUrgent— the last one blocks until the user responds - A Kanban ticket that shows which agent is working on it, which git worktree it's in, and a link to open the agent's terminal
The askQuestion tools are the key mechanism: when an agent hits an ambiguity or needs a decision, it calls the tool instead of guessing. VibeLoop routes the question to the user as a notification. askQuestionUrgent blocks the agent until the user answers, keeping the agent on the right track without requiring the user to monitor a terminal.
Architecture
- GraphQL API (graphql-yoga) for all data — projects, tickets, agent sessions, worktrees
- SSE for streaming terminal output from server to browser
- Drizzle ORM with SQLite by default, PostgreSQL supported
- Zustand for client state
Sandboxed agent execution
Agents are deployed to AWS VMs on demand. Each agent process runs inside minivisor (KVM hypervisor sandbox) with ptracer providing a full syscall-level audit trail — every file read, write, and subprocess is recorded. This means you can replay exactly what an agent did, enforce filesystem policy (e.g. only allow writes inside the worktree), and detect if an agent tries to do something unexpected.