juggler-ai/juggler

What it solves

Juggler provides a visual workbench for AI coding agents, moving away from simple scrolling chat transcripts to a structured, navigable environment. It solves the problem of "black box" agent execution by making the model's context, tool calls, and internal transactions fully visible and editable, while allowing the agent to run on a remote server where the code actually resides.

How it works

Juggler operates as a client-server architecture. A headless Go server runs on the machine containing the code, while a native desktop app or browser acts as a synchronized client. Instead of a linear chat, conversations are organized as persistent trees with nested threads. It uses a JavaScript-based extension system to define tools (like bash or file read/write), strategies for LLM loops, and UI components. It also integrates with the Model Context Protocol (MCP) to allow developers to test and debug MCP servers by inspecting the exact schemas and arguments passed between the model and the server.

Who it’s for

It is designed for developers who use AI coding agents and want granular control over the agent's context, the ability to debug model transactions, and a flexible way to manage complex coding tasks through branching conversations.

Highlights

• Tree-based Conversations: Uses Miller columns to navigate nested threads, allowing users to branch off for research or tangents without polluting the main chat.
• Context Surgery: Allows users to fold history into threads, move or copy items, and edit the prompt identity to manually refine what the model sees.
• Transaction Inspection: Provides full visibility into every LLM request, including system prompts, tool definitions, token usage, and stop reasons.
• Remote Execution: The server runs where the code is, allowing multiple clients (desktop or mobile) to connect to a single live session.
• Extensible Architecture: Supports custom JavaScript extensions for tools, commands, and UI cards, and serves as a workbench for MCP server debugging.