Super Studio

Visual agent orchestrator and skill catalog for managing 50 agents at once.

What is this

Super Studio is a control surface for concurrent AI coding agents. Instead of one agent in one terminal, you run many at once as live nodes on a graph: start them, watch them think and act in real time, approve or deny the actions they want to take, chat with any of them, and review what they produce.

This is a real tool, not a mock. A small control server on your Mac drives actual Claude Code sessions through the Claude Agent SDK, mirrors any agent sessions you started from a terminal, and can fan work out to cloud agents running on GitLab CI. The interface is a web app: run the engine and open it in any browser, or install the optional Mac app that wraps the same engine and UI.

Why it exists

Coding agents are most useful in numbers, but a terminal only shows you one at a time. The moment you run three or four, you lose the thread: which one is blocked waiting on you, which one is burning tokens, which one just edited a file you care about. Super Studio puts the whole squad in one place. Every agent is a node, every action that needs a human decision lands in a single queue, and every result shows up somewhere you can read and annotate. You supervise the fleet instead of babysitting each tab.

Who it is for

Developers and teams who lean on coding agents and want to run several at once. If you already live in Claude Code, and occasionally reach for Codex, Cursor, Gemini, or Copilot, and find yourself flipping between terminals to keep track, this is the cockpit for that.

Features

• Visual canvas of live agents. Every agent is a node showing live status, model, and token/cost counters, with reasoning and tool calls streaming in. Also renders as Kanban and List.
• Decision stack. Pending approvals and questions across the whole squad queue as one card stack; the top card pans to its node so you Allow, Allow all, or Deny right there.
• Multi-engine, local and cloud. Claude Code runs locally via the Agent SDK; cloud agents (Claude, Codex, Cursor, Gemini) run as one-shot GitLab CI jobs; terminal claude sessions mirror in as read-only nodes.
• Initiatives, artifacts, and jobs. Group agents under an Initiative, review their Artifacts with inline annotations and diffs, and track recurring scheduled Jobs on their own surface.
• Detachable mini window. Pop the decision stack into a floating window that stays above everything, so you keep approving while you work elsewhere.
• Mac app. Menu-bar app with a setup wizard, start-at-login, and an optional always-on background service.

Screenshots

Canvas	Decision stack	Side panel

Web browser or Mac app

Super Studio is a web app. A small engine (the control server) runs on your Mac and serves the entire UI over HTTP, so there are two ways to use it:

• Web browser: start the engine and open it in any browser, on the Mac itself or from another machine over the optional tunnel. Nothing to install beyond the repo.
• Mac app: a menu-bar wrapper around the same engine and UI. It starts the engine for you, injects the access code, and adds native conveniences.

The engine always runs on a Mac with Claude Code installed. This is a local tool, not a hosted service, so "web browser" means the client is a browser, not that the whole thing lives in the cloud. Everything that actually orchestrates agents is identical in both; the Mac app only adds convenience and native OS integration.

Capability	Web browser	Mac app
Canvas, agents, decision stack, approvals	✅	✅
Chat, code, preview, inspector panels	✅	✅
Local Claude Code agents (Agent SDK)	✅	✅
Cloud agents on GitLab CI	✅	✅
Terminal session mirroring	✅	✅
Jobs, Initiatives, Artifacts	✅	✅
Runs on any OS (Windows, Linux, iPad, …)	✅	macOS only
Starts the engine for you	you start it	✅
Access code	enter once per browser	injected automatically
Menu-bar tray (copy URL/code, live usage, start-at-login)	❌	✅
Native notifications (approvals, completions)	❌	✅
Detached mini window floats over every Space	browser popup	✅
Squad keeps running when you close the window	manage the engine yourself	✅

Quickstart

Requirements

• macOS with Claude Code installed and logged in (claude --version works). Local agents run under whatever auth your Claude Code uses.
• Node.js 20 or newer, and Homebrew.
• Optional, for cloud agents: glab authenticated against gitlab.com (glab auth status).
• Optional, for remote access from another machine: brew install cloudflared.

Get the code

git clone https://github.com/SupermodularAI/Super-Studio.git
cd Super-Studio
npm install
npm run build     # builds the static UI the engine serves

Use it in a browser

npm run server    # starts the engine on http://localhost:4545 and serves the UI

Open http://localhost:4545 and enter the access code (find it in ~/.super-studio/config.json). That is the whole web-browser setup. To reach it from another machine, turn on the tunnel (see Remote access and security).

Set up the Mac app

The Mac app is the same engine and UI wrapped in a menu-bar shell. Launch it straight from the repo:

npm run app       # launches the Super Studio Mac app (it starts the engine for you)

To build a double-clickable, shareable app instead:

npm run app:build   # produces dist-app/Super Studio-<version>-arm64.dmg

The DMG is ad-hoc signed but not notarized, so macOS Gatekeeper blocks the first launch of a downloaded copy. Recent macOS (Sequoia and later) removed the right-click → Open bypass for ad-hoc apps, so clear the download quarantine in one command:

xattr -dr com.apple.quarantine "/Applications/Super Studio.app"

On first launch the app walks you through a setup wizard: it checks your environment (Claude Code required, glab and cloudflared optional), picks a default working directory for new agents, and lets you leave remote access off (the default).

Develop the UI

npm run server   # control server on :4545 (serves the built UI)
npm run dev       # UI dev server on :3010 (talks to :4545)
npm run build     # rebuild the static UI the server serves

Full setup details, configuration keys, and the optional always-on background service are in SETUP.md.

How it works

any browser ──https──▶ Cloudflare tunnel ──▶ server (Mac :4545)
                                              ├─ access-code gate (REST + WS)
                                              ├─ serves the static UI (out/)
                                              ├─ Agent SDK ──▶ N local Claude sessions
                                              │    └─ canUseTool -> approvals on the canvas
                                              ├─ watcher ──▶ ~/.claude/projects (terminal sessions)
                                              └─ glab ──▶ GitLab CI pipelines (cloud agents)

• server/ is an Express and WebSocket control server (npm run server).
  • local-agent.ts runs Agent SDK sessions: streaming multi-turn input, permission and question round-trips, a subagent tree built from Task tool calls, interrupt, and token and cost tracking.
  • watcher.ts mirrors sessions started in any terminal as read-only nodes.
  • gitlab-agent.ts launches cloud agents through a GitLab CI pipeline.
  • tunnel.ts manages the optional Cloudflare quick tunnel.
• app/, components/, and lib/ are the Next.js UI, statically exported and served by the server.
• scripts/install-launch-agent.sh generates and installs the optional always-on background service.

Why real local sessions rather than the raw model API? Because a local session is a real claude subprocess with your filesystem, repos, shells, and existing MCP setup, and the SDK gives a clean approve/deny callback to pause on. See STUDY.md for the full reasoning, and docs/MULTI-ENGINE-CLOUD.md for how the multi-engine launch path works.

Remote access and security

Remote access is off by default. Super Studio is built to run locally, one install per person, not as a hosted service.

• Everything (UI, API, and WebSocket) sits behind an access code. You enter it once per browser; the Mac app injects it for you.
• Turning on the tunnel opens an anonymous Cloudflare quick tunnel that exposes your engine, including agent tool execution, to anyone who has the URL and the access code. Only enable it if you understand that, and treat the code like a password to your Mac.
• New agents can only work inside directories you list as workspace roots (your home directory by default). A working directory outside them is rejected before an agent is created.
• Local agents use Claude Code's normal permission system. Nothing is auto-approved unless you click Allow all.

Rotate the access code any time by editing ~/.super-studio/config.json and restarting the engine. The code is never written to the log file.

License

Licensed under the Apache License, Version 2.0.

Copyright 2026 supermodular.ai GmbH. See NOTICE for third-party attributions.