Parallel Code — Summary

Parallel Code is an open-source Electron (SolidJS) desktop application for dispatching multiple AI coding agents (Claude Code, Codex, Gemini, Copilot CLI) simultaneously in isolated git worktrees, with a built-in diff reviewer, AI Arena head-to-head race mode, Docker-based sandboxing via project-specific Dockerfiles, per-task step tracking (.claude/steps.json), and a QR-code mobile monitor.

Problem it solved: "Ten agents. Ten branches. One afternoon." — the same serial bottleneck as other parallel-worktree apps, but Parallel Code differentiates with AI Arena mode (race agents against each other for the same task), Docker sandboxing (drop a .parallel-code/Dockerfile to run each task in an isolated container), coverage radar (per-file test coverage badges), and steps tracking that gives an "engineering-manager-style timeline" of agent progress.

Distinctive traits: (1) AI Arena mode — race two or more agents on the same task and compare outputs; (2) Docker sandboxing per task via project-level Dockerfiles; (3) steps.json tracking written to .claude/steps.json for each task's progress timeline.

Target audience: Individual developers wanting maximum parallelism (up to 10 agents suggested), teams comparing AI tool outputs, and security-conscious users who want Docker isolation.

differs_from_seeds: Most similar to Crystal (Electron, worktrees, no methodology injection) but adds significant new capabilities: AI Arena mode (explicit parallel comparison UX), Docker isolation (the only framework in this batch supporting container isolation per task), and per-file coverage badges. Unlike superpowers (skills methodology) or claude-flow (MCP toolserver), Parallel Code is purely a session dispatch UI with no behavioral modification of agents.

Parallel Code — Overview

Origin

Created by johannesjo. MIT license. Version 1.9.0. Actively maintained as of 2026-05-23.

Philosophy

"Ten agents. Ten branches. One afternoon." "Dispatch AI coding agents in parallel, each in its own worktree. Review the diffs, merge the wins, toss the rest." "Free and open source — no extra subscription required."

Core beliefs:

1. Use the AI tools you already trust (Claude Code, Codex, Gemini, Copilot) from one interface
2. Every change isolated and reviewable — each task gets its own git branch and worktree automatically
3. Zero subscriptions: no extra platform fee beyond what you pay for the AI tools themselves

Key differentiator framing (from README)

Technology bet

Electron + SolidJS (not React) — a less common but performant frontend framework choice. This distinguishes Parallel Code from Crystal (React) and RunMaestro (React).

Target users

Developers who want maximum parallel throughput with strong isolation, including those who need Docker sandboxing for security or reproducibility.

Parallel Code — Architecture

Distribution

• Electron desktop app (macOS via .dmg universal; Linux via .AppImage or .deb)
• No Windows support (explicitly macOS and Linux)
• Download from GitHub releases or build from source

Install

npm install
npm run dev
# or download .dmg/.AppImage from releases

Directory structure

parallel-code/
├── src/               # SolidJS frontend + Electron renderer
├── electron/          # Electron main process
├── docker/            # Docker configuration templates
├── scripts/           # Build scripts
├── openspec/          # OpenSpec integration (detected in repo)
├── screens/           # Screenshots and demo GIFs
└── .vscode/           # VSCode configuration

Key project-level config

.parallel-code/Dockerfile — drop this in your project repo to enable Docker sandboxing for all tasks.

Technology stack

Required runtime

• Node.js v18+
• Git
• Docker Desktop (optional, for sandboxing)
• At least one AI CLI: Claude Code, Codex, Gemini CLI, Copilot CLI

Target AI tools

• Claude Code
• Codex CLI
• Gemini CLI
• Copilot CLI

Parallel Code — Components

UI Panels

AI Arena mode

Run multiple agents on the same task simultaneously and compare outputs side-by-side. This is Parallel Code's signature differentiator.

Docker sandboxing

Drop .parallel-code/Dockerfile in project root → each task runs inside a Docker container. Provides true isolation beyond git worktrees.

Mobile monitor

Scan a QR code → watch agent progress over Wi-Fi or Tailscale from phone/tablet.

Step tracking

Parallel Code writes agent progress to .claude/steps.json — a structured timeline that shows what the agent is doing.

Direct mode

Work on the main branch without worktree isolation (for repos without git or for direct editing).

Existing worktree import

Import already-created worktrees into Parallel Code.

No platform-level prompt templates or skills

Like Crystal and its-maestro-baby, Parallel Code injects no behavioral methodology. All AI behavior comes from the user's existing CLI configuration.

Parallel Code — Uniqueness & Positioning

differs_from_seeds

Parallel Code belongs to the same "parallel worktree manager desktop app" cluster as Crystal, RunMaestro, and its-maestro-baby, but adds two capabilities absent from all 11 seeds and all other frameworks in this batch: (1) Docker container isolation per task (.parallel-code/Dockerfile drop-in) — the only framework in this entire corpus supporting container-level sandboxing as a first-class feature; (2) AI Arena mode — a structured head-to-head race between multiple agents on the same task. Neither superpowers (skills methodology), claude-flow (MCP hive-mind), nor any other seed provides these. The openspec/ directory in the repo suggests awareness of OpenSpec, but the integration is not documented publicly.

Positioning

• "Free, open source, no extra subscription required" — explicit positioning against paid orchestration platforms
• SolidJS (not React) signals a performance-conscious design choice
• Docker support appeals to security/compliance users who need more than git worktree isolation
• Coverage radar and PR CI watching suggest appeal to teams with existing CI/CD pipelines

Observable failure modes

• macOS + Linux only (no Windows support)
• Steps tracking depends on .claude/steps.json being written — unclear whether this requires explicit agent configuration
• No automated merge strategy: user must review and merge manually
• No unattended/Auto Run mode: requires active user to create and monitor tasks

Inspired by

• Crystal (predecessor in the worktree manager space)
• Kilo Code, Roo Code (VS Code extensions cited as comparison)

Competitors named in README

• Multiple terminal windows / tmux
• VS Code extensions (Kilo Code, Roo Code)
• Running agents sequentially

Cross-references

• Uses OpenSpec (openspec directory detected) — possible integration with the spec-kit/openspec seed pattern