Back to Case Studies
Case Study

DevboardAI

A native macOS coding agent orchestrator that turns plain-English ideas into AI sprints and ships them across parallel agents — Claude Code, Codex, and Kimi — while you sleep.

Industry:Developer Tools / Desktop Software
Model:$74 one-time, local-first
devboardai.com

In one paragraph

DevboardAI is a native macOS AI coding agent orchestrator built by Sophylabs that gives developers a local-first Kanban dashboard for managing Claude Code, OpenAI Codex, and Kimi. It turns a plain-English description into a full sprint — tasks, dependencies, and validation criteria — then runs the backlog autonomously across parallel coding agents in isolated git worktrees, retrying failures with QA feedback. The product ships as a notarized Electron app for Apple Silicon and Intel, with a one-time $74 license instead of stacking $20–200/mo subscriptions.

The Challenge

Every AI coding tool on the market is an assistant: it waits for a prompt, suggests code, and leaves the orchestration to the human. The developer becomes the loop — copy a snippet, paste it into a file, run it, paste the error back into chat, repeat. Side projects pile up because every feature requires hours of supervision.

The DevboardAI founder came to Sophylabs with a clear thesis: agent CLIs like Claude Code, Codex, and Kimi are finally capable of writing real code against a real repo — they just need a builder layer on top to handle planning, task delegation, parallelism, retries, and review. Build that layer as a local-first Mac app and developers stop babysitting AI and start shipping.

The platform had to:

  • Orchestrating multiple long-running agent CLIs in parallel without one crash taking down the whole sprint.
  • Isolating agent edits so parallel coding agents can touch the same repository without race conditions or merge chaos.
  • Keeping the entire app — including agent processes, terminal sessions, and run history — local-first on the user's Mac with zero cloud middleman.
  • Designing a retry loop that turns model failures into useful QA feedback for the next attempt, instead of just looping on the same error.
  • Shipping a notarized, auto-updating native macOS app on Apple Silicon and Intel that feels like a desktop tool, not a wrapped web page.

Our Solution

We designed DevboardAI as a desktop-first orchestration engine. The UI is a Kanban board; the backend is a Node.js process inside Electron that manages a pool of agent CLIs, each running in its own git worktree. A scheduler walks the dependency graph of the active sprint, picks the right model per task, streams output to the terminal, and feeds QA failures back into the next attempt — all without leaving the user's machine.

Key technical implementations

Kanban Board for AI Agents

A five-column board — Backlog, In Progress, QA, Done, Failed — with drag-and-drop. Each card shows priority, assigned AI model, and live execution status. The single dashboard for managing parallel coding agents.

AI Sprint Generation

Type what you want to build in plain English. DevboardAI generates a full sprint with tasks, dependencies, story points, and validation criteria — ready to import in one click.

The Orchestrator

Runs every task in the backlog in dependency order, picks the right agent for each task, and retries failures with QA feedback fed back into the next attempt. Walk away and come back to a finished sprint.

Multi-Agent Execution

Mix and match Claude Code, OpenAI Codex, and Kimi per task. Value Mode auto-routes between Haiku, Sonnet, and Opus based on task complexity to keep cost down without sacrificing accuracy.

Git Worktree Isolation

Each agent runs in its own git worktree, so parallel coding agents can edit the same repo without stepping on each other. Merge or discard per task from inline diffs.

Built-in Developer Tools

Multi-tab terminal, inline Git diffs, file explorer for context attachment, and full project export — all without leaving DevboardAI. Native macOS notifications and dock badges when the sprint is done.

Architecture Highlights

Local-first by default

Project files, sprints, run history, and agent output all live in SQLite on the user's Mac. The only network calls are the agent CLIs talking to their own provider APIs. No SaaS backend to leak, breach, or shut down.

Agent-agnostic execution layer

A thin adapter wraps each CLI — Claude Code, OpenAI Codex, Kimi — behind a common interface for spawn, prompt, stream, and cancel. Adding a new agent is a single adapter, not a refactor.

Worktree-based parallelism

Every running task gets a dedicated git worktree, so multiple agents can edit the same repo at the same time without conflicts. Merges and discards happen per-task from inline diffs.

QA retry loop

Failed tasks move to QA and the error output is fed into the next prompt as structured context. The orchestrator retries with exponential backoff and falls back to a stronger model before giving up.

Value Mode model routing

Cheap models for summarization, labeling, and simple edits. Flagship models for risk-scoring big changes. The orchestrator picks per task so the user doesn't have to think about pricing tiers.

Native macOS distribution

Notarized Electron binary for Apple Silicon and Intel, Sparkle auto-updates, native notifications and dock badges. Feels like a desktop tool, not a wrapped web page.

Technical Stack

Electron
Desktop Shell
TypeScript
Language
React 19
UI
Tailwind CSS
UI
Node.js
Runtime
SQLite
Local Storage
Git Worktrees
Isolation
node-pty
Terminal
Claude Code CLI
AI Agent
OpenAI Codex CLI
AI Agent
Kimi CLI
AI Agent
Next.js 16
Marketing Site
Stripe
Payments
macOS Notarization
Distribution
Sparkle
Auto-Update
PostHog
Analytics

Results

$74
Lifetime, Not $20–200/mo

One-time payment instead of stacking Cursor, Devin, and Claude Max subscriptions. Breaks even against a single $20/mo tool in roughly 23 days.

0
Bytes Leave the Machine

Source code and repository context never leave the user's Mac. Agents shell out locally; only the agent CLI talks to its own provider API.

3+
Agent Providers Supported

Claude Code, OpenAI Codex, and Kimi out of the box, with Value Mode routing each task to the cheapest model that can finish it.

1-click
Sprint to Shipped Code

Describe a feature in plain English, generate the full sprint, hit run on the orchestrator, and come back to merged work in your repo.

"Sophylabs treated DevboardAI like their own product, not a delivery contract. They pushed back when scope drifted, made the right architecture calls early — the worktree isolation and event-driven orchestrator were both their suggestions — and shipped a desktop app that actually feels native. The orchestrator alone changed how I work on side projects."
DevboardAI
Founder

Services Delivered

Product architecture
Native macOS app development
AI/LLM integration
Multi-agent orchestration engine
Git worktree isolation
Local-first data model
Marketing site & SEO
Stripe checkout & licensing
Code signing & notarization
Auto-update infrastructure

Want to build an AI product like this?

We design and ship production AI platforms for founders and operators. Book a free call and we will tell you, honestly, whether your idea is fundable as-is.