Parallel Code
Dispatch Claude Code, Codex, and Gemini coding agents in parallel, each isolated in its own git worktree — run ten branches in one afternoon and merge only the wins.
Positives: worktree isolation, project-level Dockerfile sandboxing, semgrep security rules with self-tests, gitleaks secret scanning, CI actions pinned by commit SHA. Deducted for: no documented user confirmation flow for merge/push to remote, 'Share agent auth' and Direct mode weaken isolation, weak disclosure of code Q&A being sent to external LLMs (MiniMax), no explicit rollback/undo design, and author identity resting solely on package. with no signing or provenance.
README, package., and CI are largely self-consistent with clearly layered test scripts (unit/client/pty integration). Deducted for: multiple beta xterm dependencies and external CLI requirements with unconstrained versions; error message quality is not evidenced.
Clear developer audience on macOS/Linux, multiple CLIs, Docker isolation and non-git folders supported; Antigravity Docker auth limitation is explicitly disclosed (good boundary reporting). Deducted for: trigger conditions of features are vague and no Windows support; runtime fit unverifiable from README alone.
Install notes (release/source builds), shortcut table, comparison table, screenshots, and MIT license consistent between README and LICENSE (full marks there). Deducted for: no CHANGELOG file, known limitations scattered in collapsed sections, no documented maintenance/update commitment beyond npm version patch; version 1.15.0 exists but changelog is absent.
Core value is clear (parallel agents + worktree isolation + diff review) with differentiated keyboard-first and phone monitoring. Deducted for: token/multi-agent cost implications unaddressed; effectiveness claims are static assertions without execution evidence.
CI includes a coverage floor, self-tested semgrep rules, and a pty integration test, so claims are partly traceable to scripts. Deducted for: only four files reviewed, screenshots/demos cannot be cross-verified against code, and README mixes marketing prose with factual claims without separation.
- Static review only; no execution was performed. Inspect the Electron main process and MCP server code before installing.
- Merge-to-main, push-to-remote, and Direct mode all bypass isolation — verify the target branch before acting.
- Code Q&A (MiniMax) and phone-based remote monitoring send code/data off-machine; use caution on sensitive repositories.
- Antigravity cannot authenticate under Docker isolation and must run natively; 'Share agent auth' widens credential exposure.
- Several terminal dependencies are beta builds, and operation depends on availability and versions of external CLIs.
What does this agent do, and when should you use it?
Parallel Code (GitHub: johannesjo/parallel-code) is a free, MIT-licensed desktop application built with Electron, SolidJS, and TypeScript, available for macOS and Linux. Its core mechanism is per-task git isolation: when you create a task it makes a new branch from main, sets up a git worktree, symlinks gitignored directories like node_modules, and spawns your chosen AI coding CLI inside that worktree. One interface manages Claude Code, Codex CLI, Gemini CLI, Antigravity CLI, and Copilot CLI side by side, with parallel sessions, an inline-comment diff viewer, an AI Arena head-to-head race mode, and QR-code remote monitoring from your phone. Finished work merges back to main from the sidebar. The app ships no model itself — you must install and authenticate at least one AI coding CLI for it to do anything.
On task creation, Parallel Code: 1) creates a new git branch from main; 2) sets up a git worktree so the agent works in a separate directory; 3) symlinks node_modules and other gitignored directories into the worktree; 4) spawns the AI agent there. Optionally, if a .parallel-code/Dockerfile exists in the project, tasks run inside a Docker sandbox. On the output side: a built-in diff viewer supports inline review comments and per-commit navigation; a Steps panel renders a timeline of agent progress (written to .claude/steps.); each task gets a notes panel and its own shell terminal scoped to the worktree; a PR CI status watcher sends desktop notifications when GitHub checks settle. Merging happens from the sidebar (Ctrl+Shift+M). The inline code Q&A feature uses Claude Code by default or MiniMax M2.7 (204K context), configurable in Settings.
- Solo developers juggling multiple features: hand five tasks to five agents running simultaneously in isolated branches with zero conflicts.
- Engineers comparing models or CLIs: use AI Arena mode to race Claude Code against Codex on the same task and merge the better diff.
- tmux users tired of manual worktree bookkeeping: get automatic branch/worktree isolation, a merge button, and GUI diff review.
- People stepping away from the desk: scan a QR code and watch agent progress over Wi-Fi or Tailscale from a phone.
- Security-conscious teams: run tasks inside a sandbox via a project-specific .parallel-code/Dockerfile.
- Keyboard-first power users: every action has a shortcut (Ctrl+N for new task, Ctrl+Enter to send prompt), mouse optional.
What are this agent's strengths and limitations?
- True parallel isolation: every task automatically gets its own git branch and worktree — no manual bookkeeping, zero conflicts across concurrent agents.
- Multi-CLI cockpit: Claude Code, Codex CLI, Gemini CLI, Copilot CLI (plus Antigravity CLI) driven from one interface, avoiding lock-in to a single vendor.
- Keyboard-first with remote monitoring: full shortcut coverage plus phone-based progress watching over Wi-Fi or Tailscale.
- Auditable review flow: built-in diff viewer with inline comments, per-commit navigation, and AI Arena comparison mode.
- Optional Docker sandboxing: isolate agent execution per project via a .parallel-code/Dockerfile.
- macOS and Linux only — no Windows build is offered.
- Depends on external AI CLIs: the app is free but you must install and authenticate Claude Code, Codex CLI, etc., and pay their model costs separately.
- Docker-sandboxed Antigravity tasks cannot authenticate (keyring dependency) and must run natively — a documented constraint.
- Direct mode works on the main branch without worktree isolation; non-git folders are supported but lose the isolation guarantees.
- Building from source requires Node.js v18+ and npm; as an Electron desktop app it carries the usual resource footprint.
How do you install or deploy this agent?
Option 1: download from the GitHub releases page (github.com/johannesjo/parallel-code/releases/latest) — macOS: .dmg (universal); Linux: .AppImage or .deb. Option 2: build from source:
git clone https://github.com/johannesjo/parallel-code.git
cd parallel-code
npm install
npm run devPrerequisites: Node.js v18+. You must also install and authenticate at least one AI coding CLI: Claude Code, Codex CLI, Gemini CLI, Antigravity CLI, or Copilot CLI.
How do you use this agent?
1) Open Parallel Code and point it at a git repo. 2) Create a task (Ctrl+N) and pick an agent — the app auto-creates the branch and worktree and launches the agent inside it. 3) Review changes in the diff viewer with inline comments, or write prompts in the task notes panel and send them to the agent. 4) When satisfied, merge the task back to main with Ctrl+Shift+M or push to remote with Ctrl+Shift+P. 5) Use Ctrl+Shift+F for focus mode and Ctrl+/ to list all shortcuts. For Docker sandboxing, drop a .parallel-code/Dockerfile at the project root; note that Antigravity CLI must run natively (not in Docker) because it authenticates via the OS keyring.
How does this agent compare with similar options?
The README's own comparison: multiple terminal windows/tmux lack a GUI and automatic git isolation — you manage worktrees, branches, and merges by hand; VS Code extensions (e.g., Kilo Code, Roo Code) are tied to VS Code and lack true parallel worktree isolation between agents; running agents sequentially blocks your workflow on one task at a time. Parallel Code targets exactly those gaps with a GUI, automatic isolation, and genuine parallelism.