Juggler

// Overview

Juggler is an open-source coding agent with a graphical workbench instead of a terminal transcript, built by Jules Storer, the developer behind JUCE, Tracktion and Cmajor. Its premise is that using a coding agent means reading and editing large blocks of text under a firehose of information, and that a scrolling chat log is a poor surface for that. So a Juggler session is a tree laid out in Finder-style Miller columns: tool calls, approvals, item properties and raw context sit side by side and you navigate rather than scroll.

The session is a Yjs document on disk, not a transcript. You can create sub-threads, drill into them, backtrack, compare branches and edit what is there. Because it is a document, quitting and relaunching resumes every conversation exactly where it was — including states like an agent waiting for your approval, so the dialog is still there after a restart. Everything is inspectable: select the system prompt to see exactly what the model was told and which tools it may call, or click the footer's token count to read back what a past turn actually sent.

It is extensible in JavaScript — context items, slash commands, LLM loop strategies and their UIs are all extensions you can inspect, fork or replace, and MCP servers and skills plug into the same model so their tools arrive as inspectable, approvable items. The app runs locally, remotely, or both at once: a headless `juggler` server can host a long-lived session that the native desktop app and browser tabs attach to simultaneously. Models come from your existing subscriptions or keys — the README names Claude Code (CLI or API), OpenAI (Codex plan or API), GitHub Copilot, Gemini, Mistral, Z.ai, Ollama, OpenRouter and DeepSeek. It is free, needs no signup, and is licensed AGPL-3.0.

// What it does

• Miller-column workbench: tool calls, approvals, item properties and raw context laid out for navigation instead of scrolling
• Sessions are trees, not transcripts — create sub-threads, drill down, backtrack, compare and edit
• Persistent and stateful: a session is a Yjs document on disk, so a restart resumes even a pending approval dialog
• Full inspectability — read the exact system prompt and tool list, and replay what any past turn actually sent
• JavaScript extensions for context items, slash commands, LLM loop strategies and their UIs; MCP servers and skills use the same model
• Local, remote or both: a headless server plus desktop and browser clients can attach to one session at the same time
• Bring your own model — Claude Code, OpenAI/Codex, GitHub Copilot, Gemini, Mistral, Z.ai, Ollama, OpenRouter, DeepSeek and others

// Getting started

Every download contains the same two parts: the native desktop app, and `juggler`, a headless command-line server for long-lived, remote or network-accessible sessions. No signup is required.

open https://github.com/juggler-ai/juggler/releases

juggler

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Staying hands-on with what an agent is doing to a codebase, rather than reading a scrolling log after the fact
• Branching a long agent session into sub-threads and comparing the results
• Keeping a long-lived agent session on a remote machine and attaching to it from a laptop browser
• Auditing exactly what a past turn sent to the model, token count included

// How Juggler compares

Juggler alongside other open-source autonomous coding agents tools AI/TLDR tracks, ranked by GitHub stars.