rubberduckhq 0.3.0

Local-first orchestrator for AI coding agents — one window over your fleet

🦆 Rubberduck

One window over every AI coding agent you run.

Launch Claude Code, Codex, or Copilot into isolated git worktrees, watch them work side by side, fork a session into a tree of attempts, and answer their prompts — all from one local dashboard. Your code and keys never leave your machine.

Install · Features · Dashboard · Using it

Works with Claude Code · Codex · GitHub Copilot CLI · any CLI agent

---

Install

pipx install rubberduckhq

Requires Python 3.11+. Then, once:

rubberduck install-hooks --global   # wire your agent to report to Rubberduck
rubberduck serve                    # run the server + dashboard at :4200

Open http://localhost:4200 and start your agent — sessions appear on their own.

What it does

🪟 One window over every agent

Stop juggling five terminals. Every running session shows up in one dashboard — what each is doing, which one needs you, and how they relate.

🌿 Isolated git worktrees

Run several agents on the same repo without collisions. Each session can get its own git worktree on its own branch; merge any branch back with a normal git merge. Or run in place when you don't want a branch — your choice per session.

⑂ Fork into a tree of attempts

Fork a running session — the code (a new worktree branch) or just the conversation — and follow the whole lineage as a tree. Try three approaches to the same problem side by side.

🔔 Human in the loop (HITL)

Live state for every session: busy, idle, or waiting on a permission request. The Needs human panel surfaces the sessions waiting on you — answer their prompts from the dashboard for sessions Rubberduck launched, without switching terminals.

📜 Durable history

Every session is kept with its stated intention and an outcome summary, plus checkpoints of prompts, files, and tools. Catch up later without re-reading transcripts.

🔒 Local-first and private

The server runs on 127.0.0.1, gated by a per-install secret so a web page you visit can't drive it. Rubberduck never sees your code or your API keys.

The dashboard

http://localhost:4200 is three columns — your agents, what needs you, and a live pulse of activity (pictured at the top of this page).

• Agents — every session as a row, forks nested under their parent. A badge marks each watched (you started it) or launched (Rubberduck started it and can drive it); a ⎇ glyph marks the ones on a git branch. Click a row for its detail drawer — timeline, diff, live output, checkpoints, and inline notes.
• Needs human — the human-in-the-loop (HITL) panel: sessions waiting on a permission request, with inline Approve/Deny for the ones Rubberduck can reach.
• Pulse — a live ticker of the latest action across every agent.

Using it

Two commands: install-hooks runs once per machine (it persists across reboots), and serve runs whenever you want Rubberduck active.

Claude Code

rubberduck install-hooks --global
rubberduck serve

Start claude as usual — sessions appear automatically. The richest integration: hook events plus the JSONL transcript.

Codex

rubberduck install-hooks --agent codex --global
rubberduck serve

One extra step: Codex won't run a hook until you trust it. Start codex, run /hooks, and trust the Rubberduck hook (re-trust if Rubberduck later updates the hook script). Use --global — Codex's repo-local hooks are unreliable upstream.

GitHub Copilot CLI

rubberduck install-hooks --agent copilot --global
rubberduck serve

Start copilot as usual — sessions appear automatically, no extra steps.

Bring your own agent

Any CLI agent works through the launch path: click New session in the dashboard, point it at a command and a folder. Rubberduck spawns it in a terminal tab and supervises the process directly — no hooks needed. State is read from the agent's terminal output, so it's coarser than the hook integrations above, but it works for anything.

Two kinds of session

• Watched — you start the agent in your own terminal; its hooks report to Rubberduck. Rubberduck observes but doesn't own the process, so you drive it and answer its prompts in your terminal.
• Launched — you click New session; Rubberduck opens a terminal tab, runs the agent, and owns the process — so you can drive it and answer permission requests from the dashboard.

State lives in ~/.rubberduck/ (override with RUBBERDUCK_HOME).