HormesisWorkspace

The cockpit for people who live in the terminal. Your agents don't just run in it; they operate it and build tools into it.

Free, open-source (Apache-2.0) desktop workspace for running many coding agents in real terminals. Your own Claude subscription; nothing resold.

// every pane, every project, every restart

~/billybox $ claude --resume
~/roboscop $ claude — fixing #12
[tool] competitor-table · live
~/logrider $ zsh
radar · 7 agents on host
~/utanga $ claude — tests passing
demo — 60s, coming with launch

Three Pillars

01 / agent-operable

Agents operate — and furnish — the workspace

The workspace is itself an MCP server. Agents open, split and resize panes, spawn other agents, queue prompts into them — and author live tool panes from JSON (tables, trackers, dashboards) whose buttons route back to the authoring agent. Named tools persist and accumulate into your personal Tools shelf.

02 / machine-wide radar

See every agent on your machine

A machine-wide radar sees every Claude process on the host — even ones started in iTerm, by scripts, or by other apps. Read-only by default; attach and take over on demand. One pane of glass for everything running.

03 / state survives

Kill it. Nothing is lost.

Every pane auto-resumes its exact conversation across restarts — session IDs are assigned up front, so resume never collides. Scrollback survives force-quits and sleep/wake: journaled to disk, alt-screen aware.

An Upgrade in Place, Not a Migration

The CLI is the product. Claude Code's growth was earned overwhelmingly in the terminal — and that's where this workspace lives. Every pane is a real PTY: your shell, your dotfiles, the actual claude CLI, same TUI, same keybindings.

GUI-first tools ask you to migrate: a chat transcript in the center, satellite panes around it, one vendor, one repo at a time. Adopting HormesisWorkspace is an upgrade in place — the console workflow you already have, plus persistence, radar, portfolio scope and agent-operability. Shells, ssh, codex panes, any CLI agent in the same tree. Zero retraining.

Works With Claude Code Desktop

The official desktop app does parallel sessions well. This is a different theory of the screen: Claude Code Desktop's primary object is a chat session; ours is the terminal itself — your shell, your TUI, zero retraining, N projects side by side. The radar also sees sessions the official app can't (anything on the host), and the workspace itself is agent-operable via MCP, which has no first-party equivalent. Run both; the radar will see it.

Install

  • macOSDMG from GitHub Releases (Apple Silicon + Intel)
  • WindowsInstaller from GitHub Releases (x64)
  • LinuxAppImage from GitHub Releases

How It Compares

HormesisWorkspaceVibe KanbanConductortmux + scripts
Primary objectThe terminal itself — real PTYsKanban board of agent tasksChat sessions per repoRaw panes, no agent layer
Agents operate the workspaceYes — workspace is an MCP serverNoNoNo
ScopeWhole portfolio, any CLI agentMulti-CLI, task-scopedmacOS-only, repo-scopedAnything, manually
StatusActively developed, Apache-2.0SunsettingFree, closedEvergreen

Full comparison: vs Vibe Kanban