coder/xum

What it solves

Xum is a coding agent multiplexer that allows developers to manage and run multiple AI coding agents across different isolated workspaces. It solves the problem of managing multiple agentic workflows up front, providing a central view of git divergence and cost tracking across various models.

How it works

Xum provides a custom agent loop with features like Plan/Exec mode and vim inputs. It supports multiple runtimes for isolated execution, including local directories, git worktrees, and remote servers via SSH. It integrates with various LLMs via Ollama (for local models) and OpenRouter (for cloud models), and offers a VS Code extension for seamless transition into agent workspaces.

Who it’s for

Software engineers who want to use multiple AI coding agents simultaneously and manage their changes across different git branches or remote environments without interfering with their main development flow.

Highlights

• Isolated Workspaces: Support for local, git worktree, and remote SSH execution.
• Multi-model Support: Compatible with models from Anthropic, Grok, OpenAI, and local LLMs via Ollama.
• Model Multiplexing: Ability to manage a suite of agents with dedicated UI and keybinds.
• Rich Output: Supports Mermaid diagrams and LaTeX for complex proposals.
• VS Code Integration: Direct jump into Xum workspaces from the editor.
• Context Management: Features like opportunistic compaction and mode prompts to manage agent context.