Skip to main content
Uzi provides a suite of commands for managing AI coding agents across multiple isolated environments. Each agent runs in its own tmux session with a dedicated git worktree.

Available Commands

prompt

Launch agents with a prompt in isolated environments

auto

Automatically manage agent sessions and handle prompts

ls

List all active agent sessions with status

broadcast

Send messages to all active agents simultaneously

run

Execute commands across all agent sessions

checkpoint

Merge agent changes back to main worktree

kill

Terminate agent sessions and clean up resources

reset

Delete all Uzi data and start fresh

Quick Reference

CommandAliasDescription
promptpRun prompt with specified agents
autoaAuto-manage agent sessions
lslList active sessions
broadcastbSend message to all agents
runrRun command in all sessions
checkpointcRebase agent changes
killkDelete agent session
reset-Reset all Uzi data

Common Workflows

Starting Agents

# Launch a single Claude agent
uzi prompt "implement user authentication"

# Launch multiple agents with different models
uzi prompt --agents=claude:2,codex:1 "fix the login bug"

Monitoring Progress

# List all active agents
uzi ls

# Watch mode with live updates
uzi ls -w

# Auto-manage prompts (hands-free operation)
uzi auto

Managing Changes

# Checkpoint an agent's work
uzi checkpoint agent-name "feat: add authentication"

# Terminate an agent
uzi kill agent-name

# Clean up all agents
uzi kill all

Global Concepts

Agent Sessions

Each agent runs in an isolated environment consisting of:
  • Tmux session for terminal interaction
  • Git worktree for isolated file changes
  • Git branch for version control
  • Dev server (optional) on a dedicated port

Session Naming

Sessions follow the pattern: agent-{project}-{hash}-{agent-name} Example: agent-uzi-a3f2b1c-thoughtful-tesla

State Management

Uzi tracks all sessions in ~/.local/share/uzi/state.json with:
  • Prompt text
  • Branch name
  • Session name
  • Worktree path
  • Model name
  • Dev server port (if configured)

Build docs developers (and LLMs) love