← All projects
AIOpen SourceElectronTypeScript

Open source

Agent Workbench

A desktop control plane for running multiple AI agents — Claude, Codex, Gemini, shell, and Python REPL — from a single window. PTY-backed terminals, voice input, and an operator panel that steers agents via tool calls.

5+

Concurrent agent types (Claude, Codex, Gemini, shell, Python)

Local-first

State and logs persist at ~/.agent-workbench — no cloud

Voice

Push-to-talk operator panel — hold Space anywhere to dictate

The challenge

What needed to change

Running multiple AI agents from separate terminals creates context-switching overhead and no shared view of what each agent is doing. There was no single interface for steering multiple AI CLIs, capturing their output, and issuing commands via voice or text.

The approach

How I built the system

Built on Electron + Next.js with a Node PTY sidecar. Each agent is a real PTY child process with full terminal I/O. An operator panel lets you spawn, kill, and direct agents via tool calls. Push-to-talk voice input goes straight to the operator. State persists locally — nothing leaves your machine beyond the AI APIs you're already signed in to.

The outcome

What exists now

  • → Multi-agent terminal control plane with shared operator visibility.
  • → Voice-to-agent dispatch from any context in the app.
  • → Session persistence — exited agents stay visible for output review.