The Problem
Parallel AI coding agent workflows currently require juggling multiple separate CLIs (Claude Code, Codex CLI, etc.) without integrated workspace isolation, making it difficult to verify agent actions, maintain code integrity, or coordinate across multiple simultaneous tasks. Teams either accept unverifiable "black box" coding or manually manage isolated environments, creating overhead that slows delivery.
What This Does
Manaflow is an Electron-based desktop application that spawns multiple coding agent CLIs in parallel, each running in its own sandboxed VS Code workspace within Docker containers. The apps/client/ directory contains the React/TypeScript frontend with 141 files managing agent orchestration, while .beads/config.yaml and .beads/interactions.jsonl define the system's behavioral contracts. Isolated VS Code environments are provisioned per-task, with live web preview, terminal access, and git diff view integrated directly into the UI (apps/client/src/components/ contains the component hierarchy for task panels, terminals, and diff viewers). A heatmap diff viewer automatically annotates risky code changes (apps/client/src/components/RunDiffHeatmapReviewSection.tsx). One-click PR creation gates on reviewed diffs (apps/client/src/components/delete-button.tsx, apps/client/src/components/restore-snapshot-button.tsx).
How To Use It
Setup: The application is distributed as a macOS binary (with Linux in beta per the README). The Electron app is built with Vite and React (apps/client/electron/vite.config.ts, apps/client/src/app.tsx). No lockfile exists for apps/client/package.json, meaning dependency versions may shift between installs.
Configuration: Environment variables are expected in .env.local.template at the root; the Electron main process reads these via apps/client/electron/main/electron-main-env.ts. Convex backend configuration lives in .devcontainer/docker-compose.convex.yml for cloud deployment. The .claude/settings.json and .claude/hooks/ directory configure Claude Code integration hooks.
Running it: Launch the Electron application directly from the downloaded macOS binary. No local build step is required for end users; the packaged app bundles the React runtime. For development, docker-compose.yml in .devcontainer/ spins up the full stack (Electron + Convex backend) with hot-reload.
Real-World Use
A distributed engineering team preparing a release runs three parallel agents: one refactors the payment service, another writes integration tests, and a third updates the changelog. Each agent operates in its own VS Code workspace with the codebase mounted, allowing real-time observation of edits, terminal output, and preview of the running app. After agents complete, the team reviews diffs in the built-in heatmap viewer, squashes relevant changes, and opens a single PR—all without leaving Manaflow. The VNC browser preview lets the team verify that the running dev server renders correctly across agent-induced changes.
Code Health & Issues
Low/Risk - Dependencies declared without a lockfile in apps/client/package.json, leading to non-reproducible builds across environments Med - 6 test files found; no evidence of CI pipeline execution for pull request validation beyond the GitHub Actions workflows listed (.github/workflows/) Med - No lockfile means npm install may resolve different dependency trees for different developers, risking "it works on my machine" failures Low - .env.local.template present but no .env file committed; secrets handling depends on the operator's environment setup
The repo has CI configured (17 GitHub Actions workflows), a Dockerfile for container builds, and Dockerfile.test.ts for test builds, but the missing lockfile is the most concrete risk for team adoption.
The Bottom Line
Manaflow delivers on its promise of parallel, isolated AI coding work with a polished Electron interface and real workspace visibility. It's best suited for teams that need to run multiple coding agents simultaneously and require verifiable, auditable outputs. The absence of a lockfile and sparse test coverage are the primary concerns for engineering leadership; if those are addressed, it's a credible alternative to proprietary coding agent platforms.