Open source
Agent Workbench
A desktop control plane for coordinating Claude, Codex, Gemini, shell, and Python sessions from one operator surface. PTY terminals, voice input, persistent session state, and an operator panel make multi-agent work visible enough to steer.
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
Context
Taming a sprawl of parallel agents
Running several AI coding agents at once means juggling separate terminals with no shared view of what each one is doing. Agent Workbench turns that sprawl into one controllable workspace: spawn the agents, watch their output, and direct the next move without losing the thread.
Approach
A real-terminal control plane
Built on Electron and Next.js with a Node PTY sidecar, so every agent is a real terminal child process with full I/O, not a fragile API wrapper. The operator panel handles spawn, kill, and directed tool calls; push-to-talk voice routes instructions from anywhere in the app; local state keeps the system private and inspectable.
Outcome
One cockpit for every agent
- → A shared control plane for multiple AI agents and terminal sessions.
- → Voice-to-agent dispatch from any context in the app.
- → Persistent sessions that keep exited agents visible for output review.