Juggler vs zot (2026)

Juggler and Zot both rethink the same problem (a single-user coding-agent harness you run against your own LLM) from opposite ends. Juggler, built by JUCE and Tracktion creator Julian Storer, is resolutely a GUI app: it runs as a native desktop client or a headless server reachable from a browser, syncing one live session across any number of clients via Yjs. Every conversation is an editable document rather than a log: any point can fork into a recursive sub-thread, and the whole tree is navigable and inspectable in a Miller-column view, alongside every tool call, approval and raw context message. It drives roughly six providers (Claude, OpenAI, Gemini, Ollama, OpenRouter, DeepSeek) through a swappable JavaScript plugin system. Zot, built by independent developer Patrice Eckhart, is the opposite bet: a single static Go binary with no Node runtime, Docker or account required, running as a terminal TUI, a scriptable print/JSON-streaming mode, or a headless daemon. It supports over 30 model providers (including local Ollama models) and adds a Telegram bridge that mirrors a session to your phone bidirectionally, plus "Swarm" background subagents that run in parallel on the same repository.

The branching feature both tools advertise is not quite the same thing. Zot's `/session fork` and `/session tree` genuinely branch a new session from any past message and render every session in a directory as a parent/child tree (confirmed live against zot.sh 2026-07-24), but that tree is a textual, terminal-rendered structure, not a navigable visual document. Juggler's entire product is that visual layer: branches are edited, dragged and inspected directly in a GUI, which is the exact gap an HN reviewer flagged when praising "the session tree in particular" as a genuinely differentiated take. So the honest framing is visual, editable tree navigation (Juggler) versus text-mode fork/tree plus remote and parallel-automation features a GUI app doesn't attempt (Zot).

Licensing and risk cut in opposite directions too. Juggler's core app ships AGPLv3 (a strong copyleft license, a real constraint versus the permissive MIT/Apache terms most rival open-source coding agents use), while its extension SDK is Apache-2.0 and the whole thing is free with no signup. Zot is MIT-licensed throughout, the more permissive choice, also free with no account required. But Zot carries a disclosed risk Juggler does not: its own README and HN thread confirm it reuses Anthropic's and OpenAI's first-party CLI OAuth client IDs so a Claude Pro/Max or ChatGPT Plus/Pro subscription can authenticate from a third-party tool. The README itself warns this "may be against their terms of service and may be revoked at any time" (the API-key path avoids the risk entirely). Juggler carries the more ordinary early-stage risks instead: the creator calls it alpha software, it has no ACP (Agent Client Protocol) support yet, and headless/server mode on Linux needs the `libwebkitgtk-6.0` system library as an extra deployment step.