Juggler Code Agent

Juggler is an open-source AI coding agent built for developers who want hands-on visibility and control over what an LLM is doing to their codebase. It ships as both a native desktop app and a headless command-line server, licensed under AGPLv3, with the extension SDK and bundled extensions under Apache-2.0. The project is actively maintained under the juggler-ai GitHub organization, with its latest release (v0.3.7) published in July 2026.

What It Is

Juggler is a graphical AI coding agent — not a terminal chatbot or a chat-style transcript UI. Its defining feature is a Miller column workbench (the same column-view navigation pattern used in macOS Finder) that exposes tool calls, thread structure, item properties, and raw context all at once, without burying them in collapsible chat bubbles. The session is stored as a Yjs document (a collaborative CRDT), which means it is a live, editable tree rather than a linear scroll.

The Tree-Based Session Model

Most coding agents give you a single linear transcript. Juggler structures the session as a navigable tree:

• Any point in a conversation can branch into a sub-thread
• Sub-threads can branch again, enabling parallel exploration
• You can navigate, inspect, backtrack, compare, and edit the structure directly
• The Yjs document backing the session enables multiple clients to attach and stay in sync simultaneously

This makes it practical to explore multiple approaches to a problem without losing earlier context.

Plugin Architecture

The README describes Juggler as "plugins all the way down." The core app manages document orchestration; almost everything else is a JavaScript extension:

• Context items — every item type (read-file, replace-text, bash, etc.) controls both its LLM representation and its UI appearance
• Strategies — high-level LLM loops such as plan or research are plugins, and custom loops can be added
• Commands — slash commands like /clear and /compact are plugins that manipulate the session document

The extension SDK is Apache-2.0 licensed, so closed-source extensions carry no copyleft obligation. This makes Juggler a platform for building custom orchestration workflows that need their own UI or visualizations, not just a fixed-feature agent.

Multi-Client, Local-First Architecture

Juggler's desktop app is a native window (built with Go + Wails, no Electron) that wraps a local webserver. That server is the actual session host, and any number of clients can attach to it:

• The desktop app is one client
• Browser tabs on the same machine are additional clients
• Devices on the local network can connect when LAN mode is enabled (press p in the terminal or launch with --public)
• The official binaries from juggler.studio additionally include WAN access modes not present in the open-source repository

The server is localhost-only by default. The frontend is type-checked JavaScript (JSDoc + strict static linting in CI) with no build step between source and what ships.

Model Support and Installation

Juggler connects to Claude Code (via CLI or API), OpenAI (codex plan or API), Gemini, Ollama, OpenRouter, Z.AI, Deepseek, and others. Adding new providers is described as straightforward.

Installation paths:

• macOS — .dmg download, drag to Applications; Gatekeeper bypass required on first launch
• Windows — Juggler-<version>-setup.exe installs the desktop app and juggler.exe CLI server together
• Linux — headless juggler server binary, connect via browser or desktop app

Update: v0.3.7

The latest release is v0.3.7, published July 14, 2026. The repository was created in June 2026 and has seen rapid iteration, with the default branch named develop and active pushes as recently as July 14, 2026. The project had 238 stars and 7 forks at the time of indexing, signaling early but growing community interest for a project only weeks old.