AgentStudio

AI coding agents on your Mac — describe it, and watch two AIs build it.
Two agents work side by side: one (Claude) plans & reviews, the other (Codex) writes the code. Build apps, websites & scripts in plain language — no terminal needed. Claude · Codex · DeepSeek, via App login or your own API key.

English · 简体中文 · Wiki

What is it · Use cases · Features · How it works · Comparison · Quick start · FAQ

---

What is AgentStudio?

AgentStudio is a free, open-source macOS app for AI coding — and a no-code, vibe-coding app builder that runs two AI coding agents (Claude + Codex) side by side and lets them collaborate. One lane (by default Claude) plans the work and reviews the result; the other (by default Codex) actually writes and edits the files in your project. You describe what you want in plain English or 中文, and AgentStudio turns it into working code — apps, websites, scripts, small games — without a terminal, environment setup, or reading walls of red error text.

In short, AgentStudio is a friendly GUI for Claude Code and Codex that turns "get an AI to build it for me" into describe → it builds → preview the result. Each lane can also run DeepSeek, and you connect with either an existing App login (your Claude / Codex subscription) or your own API key.

• Master lane — plan & review. The planner/reviewer agent understands your goal, breaks it into a short plan, and checks the resulting code for quality.
• Slave lane — write & execute. The builder/executor agent creates and edits real files in your project to make the idea work.

It's an AI pair-programming and no-code / vibe-coding workbench in one: developers get an orchestrated multi-agent loop; non-coders get an app builder that needs no command line.

Who is it for

AgentStudio works as a no-code, vibe-coding app builder for non-coders and as an AI pair-programming workbench for developers.

• Non-coders & makers who want to build a website, landing page, portfolio, to-do app, or small tool by just describing it.
• Developers who want a two-agent plan → build → review loop with shared memory, instead of babysitting a single terminal CLI.
• Claude Code / Codex users who'd rather drive their existing subscription from a clean desktop GUI with live preview.
• Anyone comparing AI coding tools who wants multiple models (Claude, Codex, DeepSeek) in one window, mixed and matched per task.

Features

• 🤝 Plan → code → review, fully automatic. In Dual mode the two agents collaborate on their own — no buttons to press.
• 💬 Interject anytime. Type while it's running to add or correct instructions mid-task.
• 👀 Live preview. Web pages render in an embedded browser the moment they're written.
• 🧠 Built-in memory that every model shares. A two-tier memory (your notes + auto-learned facts) is injected into all backends, so Claude, Codex, and DeepSeek stay on the same page across sessions — isolated from your machine's own CLI memory.
• 🔌 Connect your way. Per-backend App login (OAuth / subscription) or API key, with one-click connect/disconnect.
• 🎛 Pick any model. Real, auto-updating model lists per backend (Opus 4.8 / Sonnet 4.6 / Haiku … for Claude; GPT-5.5 / 5.4 … for Codex; live list for DeepSeek).
• 🌓 Polished & bilingual. Light/dark themes, English/中文 UI, conversation history with full-text search, and per-lane proxy control.
• 🔒 Local-first & private. Your prompts and API keys go straight from your Mac to the provider you chose — nothing is sent to us, and there's no tracking.
• 🍎 Native feel, no terminal. Pick a folder, say a sentence, go.

How it works

Switch modes at the top of the window:

Solo mode — chat with each agent independently (left and right are two separate conversations). Use it to drive one AI at a time.

Dual mode — give one goal and the two agents run the whole loop themselves:

You: "Build a dark-mode to-do web page"
      │
      ▼
  Claude (plan) ─▶ Codex (write files) ─▶ Claude (review the diff)
                         ▲                          │
                         └──── revise (up to 3×) ◀──┘
                                                    │
                                                    ▼
                                                 ✅ done

No clicks required end to end — and you can type an interjection at any moment to steer it.

How is AgentStudio different?

	AgentStudio	Terminal CLIs (Claude Code / Codex)	AI IDE plugins (Copilot / Cursor)
Interface	Clean desktop GUI, no terminal	Command line	Inside a code editor
Agents	Two specialized agents that plan, build & review together	One agent per session	One assistant
Models	Claude + Codex + DeepSeek, mixed per lane	One vendor per tool	Mostly one vendor
Memory	Shared across all models, global + per-project	Per-tool, separate	Editor-scoped
Built for	Non-coders and developers	Developers	Developers

AgentStudio doesn't replace these tools — it orchestrates them. App login reuses your existing Claude / Codex subscription, so you keep the plans you already pay for.

Memory

AgentStudio keeps its own memory, shared by every backend and kept separate from the machine's native CLI memory (Codex's cross-session memories is turned off for AgentStudio runs). Two tiers, each scoped global (all projects) or project (stored in <project>/.agentstudio/, so it travels with the repo and is versionable):

• Manual memory — what you write yourself, plus anything you save with a trigger phrase in chat: 记住 / 别忘了 / remember / don't forget / note that ….
• Auto memory — silently summarized from finished conversations (Codex / Claude-Code style), then injected into the next runs. Editable, with one-click Tidy (dedup/compress) and Clear.

Backends & connection

Backend	Typical role	Connect via
Claude	plan · review (or code)	App login (claude CLI / subscription) or Anthropic API key
Codex	code · execute	App login (codex CLI / ChatGPT) or OpenAI API key
DeepSeek	plan or write (text / whole-file)	DeepSeek API key

Any lane can use any backend. App login keeps your normal subscription; API-key mode injects the key into that backend only.

Quick start

Prerequisites

• macOS, Node.js 20+ and npm.
• For App login backends: the claude and/or codex CLIs installed and logged in.
• For API-key backends: an Anthropic / OpenAI / DeepSeek key (entered in the app).

Run from source

git clone https://github.com/zoyluoblue/AgentStudio.git
cd AgentStudio/studio
npm install
npm run dev        # launch in development

Build a double-clickable app (.dmg)

npm run dist       # outputs to studio/release/

Then: pick a project folder → connect a backend in each lane → choose a mode → describe what you want.

FAQ

What is AgentStudio in one sentence?

A free, open-source macOS desktop app where two AI agents — Claude (plan & review) and Codex (write code) — build software for you from a plain-language description, no terminal required.

Is AgentStudio free and open source?

Yes. AgentStudio is released under the Apache License 2.0. The app itself is free; you only pay for your own AI usage (a Claude / Codex subscription, or pay-as-you-go API keys).

Do I need to know how to code?

No. You describe what you want in plain English or 中文, pick a folder, and the agents do the building. Developers can still drop into the files at any time.

Is AgentStudio a no-code / vibe coding tool?

Yes. AgentStudio is a no-code, vibe-coding app builder: you describe an app, website, or script in plain language and two AI agents build it for you. Developers can still edit the files directly.

Do I need an API key?

Not necessarily. With App login you reuse your existing Claude or Codex subscription (via the official CLIs). Prefer pay-as-you-go? Use your own Anthropic, OpenAI, or DeepSeek API key instead. DeepSeek is API-key only.

Which AI models does it support?

Claude (Opus 4.8 / Sonnet 4.6 / Haiku …), Codex (GPT-5.x), and DeepSeek — with real, auto-updating model lists. Any lane can run any model.

Does AgentStudio work on Windows or Linux?

Not yet — AgentStudio is macOS-only today. The Electron codebase is cross-platform-friendly, so contributions toward other platforms are welcome.

Is my code and data private?

Yes. AgentStudio is local-first: your prompts and API keys go directly from your Mac to the provider you chose (Anthropic / OpenAI / DeepSeek) or your local CLI login. Nothing is sent to us, and there is no telemetry or tracking.

How is it different from Claude Code or the Codex CLI?

Those are single-agent command-line tools. AgentStudio gives them a desktop GUI, runs two agents that plan, build, and review together, shares one memory across all models, and previews the result live — no terminal.

What can I build with it?

Websites, landing pages, portfolios, to-do and note apps, dashboards, small browser games, scripts, and quick prototypes — anything you can describe.

Project structure

AgentStudio/
├─ assets/                 # logo + screenshots
├─ LICENSE                 # Apache-2.0
├─ studio/                 # Electron + React desktop app (shipping)
│  ├─ electron-builder.yml # packaging config (macOS dmg)
│  └─ src/
│     ├─ main/             # Electron main: orchestration, CLI drivers, memory, settings, auth
│     ├─ preload/          # contextBridge → window.studio API
│     ├─ renderer/         # React + Tailwind UI
│     └─ shared/           # IPC channel + type contracts
└─ app/                    # native macOS app (SwiftUI) — in development

Tech stack

Electron 42 · React 19 · TypeScript 5 · Tailwind CSS · electron-vite. The engine drives the official CLIs — claude -p (structured, read-only planning/review) and codex exec --json (real file edits, resumable sessions) — and calls DeepSeek over its OpenAI-compatible HTTP API. Code changes are reviewed via snapshot-based content diffs. A native SwiftUI macOS app is also in development under app/.

Documentation

Full guides live in the Wiki — getting started, concepts, backends & connection, the memory system, settings & proxy, and troubleshooting.

Contributing

Issues and PRs are welcome. Please run npm run typecheck (in studio/) before opening a PR, and keep changes focused. For larger features, open an issue first to discuss.

License

Apache License 2.0 © ZoyLuo.

Acknowledgements

Built on Claude Code, OpenAI Codex, and DeepSeek. UI inspired by the craft of Linear & Raycast.