JUCE creator ships Juggler, an open-source GUI coding agent built around session trees

Juggler is a new open-source AI coding agent from the creator of JUCE that rejects the standard chat-transcript interface in favor of a graphical workbench. Instead of a linear scroll you can only rewind, a session is modeled as a branching tree stored as a Yjs document: any point can spawn sub-threads, and you can drill down, backtrack, compare, and directly edit the structure. The UI borrows Finder-style Miller columns to keep tool calls, approvals, item properties, and raw context visible rather than buried in collapsible chat, targeting developers who want tight, inspectable control over what the model is doing to their code.

The architecture is plugin-first. Almost everything that composes a session — context item types like read-file and bash, high-level LLM loop strategies, and slash commands such as /clear and /compact — is a JavaScript extension you can inspect, fork, or replace. Under the hood it’s a Go backend using Wails for windowing, serving a type-checked JavaScript UI (JSDoc types, no TypeScript build step) over a local webserver, deliberately avoiding Electron. Because the app is just one client of that server, the desktop app, browser tabs, and remote machines can all attach to the same live session simultaneously, and it connects to the usual model roster: Claude Code, OpenAI, Gemini, Ollama, OpenRouter, Z.AI, Deepseek, and more.

Licensing splits the project: the application code is AGPLv3, while the extension SDK and bundled extensions are Apache-2.0, so third parties can build closed-source extensions without copyleft obligations. Networking is intentionally conservative — the server is localhost-only by default, LAN exposure is opt-in and unauthenticated, and WAN access is reserved for the official commercial binaries rather than source builds. The pitch is less “one more coding agent” and more a platform for LLM workflows that deserve their own UI instead of living as a headless script in a terminal.