Maestro: Run Multiple Claude Code Sessions in Parallel

What It Is

Maestro is an open-source orchestration tool that spawns multiple AI coding agents simultaneously in a unified interface. Available at https://github.com/its-maestro-baby/maestro, it creates a grid view where developers can run 1-12 separate Claude Code sessions (or mix in other models like Gemini) working on different tasks at the same time.

The architecture solves a fundamental problem: each AI agent gets its own Git worktree and branch, preventing the merge conflicts that would normally occur when multiple sessions modify the same codebase. Real-time status indicators display what each agent is doing, while a GitKraken-style commit graph visualizes where changes are happening across the repository. Quick action buttons handle common operations like running tests, committing changes, or pushing code without switching between sessions.

Template presets simplify setup - developers can launch configurations like “4 Claude sessions” or “3 Claude + 2 Gemini” with a single click. A plugin marketplace allows teams to attach MCP servers and custom skills to individual sessions, tailoring each agent’s capabilities to specific tasks.

Why It Matters

This tool addresses a bottleneck in AI-assisted development: the serial nature of working with coding agents. Traditional workflows involve giving an AI agent a task, waiting for completion, reviewing the output, then moving to the next feature. Maestro transforms this into parallel execution, similar to how modern CPUs handle multiple threads.

The implications are particularly significant for teams with Claude Max subscriptions or similar high-token-limit plans. The creator explicitly built Maestro to maximize token usage on client projects - instead of leaving capacity unused, developers can assign multiple agents to different features, bug fixes, or refactoring tasks simultaneously. This fundamentally changes the economics of AI-assisted development for consulting firms and agencies billing by the project rather than the hour.

The Git worktree implementation deserves attention. By giving each session its own working directory and branch, Maestro eliminates the coordination overhead that typically plagues parallel development. Developers can review each agent’s work independently, cherry-pick successful implementations, and discard failed attempts without untangling interleaved changes.

For teams already using MCP (Model Context Protocol) servers, the per-session plugin system means different agents can access different tools - one might have database access for migrations while another uses a testing framework for writing specs.

Getting Started

Clone the repository and follow the setup instructions at https://github.com/its-maestro-baby/maestro. The project requires Node.js and Git installed locally.

After installation, configure API keys for the models being used. For Claude sessions, add Anthropic API credentials to the configuration file:

{
 "providers": {
 "anthropic": {
 "apiKey": "sk-ant-..."
 }
 }
}

Launch the interface and select a template preset matching the desired configuration. The “4 Claude sessions” template creates four parallel agents, each with its own worktree. Assign tasks by clicking into each session and providing instructions as with any AI coding assistant.

Monitor progress through the status indicators and commit graph. When an agent completes its work, review the changes in its dedicated branch before merging. The quick action buttons streamline testing and committing without manual Git commands.

A detailed walkthrough is available at https://youtu.be/FVPavz78w0Y showing real-world usage patterns.

Context

Maestro occupies a unique position between single-agent coding assistants and full autonomous development frameworks. Tools like Cursor and Windsurf provide excellent single-session experiences but lack native parallel execution. Frameworks like AutoGPT and MetaGPT aim for full autonomy but often struggle with reliability on real codebases.

The Git worktree approach has limitations - it works best for feature development where tasks are naturally isolated. Refactoring that touches many files across the codebase may still require sequential work to avoid conflicts. The tool also assumes developers will actively review and merge agent output rather than automatically integrating changes.

Resource consumption scales linearly with session count. Running 12 simultaneous Claude sessions requires substantial API quota and generates significant token costs. Teams should start with 2-3 sessions to understand usage patterns before scaling up.