Superagent, explained.
Install it, point it at something, and talk. Everything below is written for someone who has never opened a terminal — and stays true for someone who lives in one.
What Superagent is
Superagent is a Mac app that gives a coding agent a place to work — and gives you a place to watch. Every project is a chat sitting next to the thing being built: a real browser on the sites you're already logged into, the files it's editing, and an iOS Simulator streamed straight into the window.
It is not an AI of its own. It runs Claude Code, on the Claude subscription you already pay for, entirely on your Mac. There is no Superagent server, no account and no API key.
You describe a thing; it gets built on your machine, on your accounts; you watch it happen in the pane next door.
Who it's for
Superagent is designed for the person who doesn't write code first, and the person who does second. That order is deliberate, and it's the opposite of how most tools in this category are built.
It doesn't get there by hiding the code — the files, the diffs and the branches are all right there when you want them. It gets there by putting the thing being built on screen next to the chat, so you can judge it the way you judge anything else: by looking at it.
You'll get on with it if…
- You'd rather see the page than read the diff.
- You want the agent on your accounts — your dashboard, your admin panel, your half-finished site — not a clean sandbox that's logged into nothing.
- You want it on your own machine, on the subscription you already have.
- You're building something on your phone and want to watch it run on a real device.
You might want something else if…
- You want a fleet. If the job is fanning one prompt across a dozen agents and racing them, a parallel-worktree tool like Orca is built for exactly that, and it runs on Windows and Linux too.
- You have no Mac and no repo. A hosted builder like Lovable or Bolt will get you a new app tonight without installing anything.
- You want errands, not software. Booking, forms, research round-ups — that's a different kind of agent.
Install it
What you need
| A Mac | Apple Silicon (M1 or later). Intel Macs aren't supported. |
|---|---|
| Claude Code | Required. Installed and signed in. Your Claude subscription is the only thing Superagent runs on — nothing extra to buy, no key to paste. Get Claude Code. |
| Xcode | Optional. Only if you want the iPhone simulator in the window. |
| An iPhone | Optional. Only if you want the companion app in your pocket. |
The steps
- Install Claude Code and sign in. Open it once and make sure it answers you. Superagent drives it, so if it isn't working there, it won't work here.
- Download Superagent. the .dmg — signed and notarized by Apple, so it opens without arguing.
- Drag it to Applications and open it. macOS will ask once whether you're sure.
- Add a project. Point it at any folder on your disk — a git repo, a website, a pile of documents. It doesn't have to be code.
- Say what you want. That's the whole interface.
New versions arrive signed and notarized in the background, with a "What's new" for each release. Restart whenever it suits you.
Or run it from source
git clone https://github.com/pungme/superagent-desktop
cd superagent-desktop/app
npm install
npm run dev
Your first five minutes
The fastest way to understand the app is to give it something with a screen.
- Add the project you're working on, or make an empty folder and add that.
- Ask for something visible. "Build me a one-page site for a coffee shop and show it to me." It will write the files, start a dev server itself, and open the page in the browser pane beside the chat.
- Ask for a change. "Make the header dark." Watch the page update in the same window — no alt-tabbing to find out whether it worked.
- Press the two-views button in the browser toolbar to see the same page as a desktop card and an iPhone frame at once, both live.
- Keep it or throw it away. On a git project, two buttons appear at the top of the chat when it has made changes. That's the whole review step.
Hold ⌥ Space, speak, let go. Transcribed on your own Mac; your voice never leaves it. The model is fetched once, the first time you use it.
Chats & branches
On a git project, every conversation automatically gets a private copy of the project on its own branch. Two chats on the same repo can edit, branch and commit side by side without touching each other — or your own checkout.
You don't have to know what any of that means to use it. It shows up as two buttons.
| Keep | Adds everything the chat did to the project as one change, named after the chat. It lands on the branch the chat was started from, even if you've switched the project since. |
|---|---|
| Throw away | Deletes all of it. Deleting a chat that still has unkept work asks you first. |
The details, if you want them
- Branches are named after the conversation. "Fix the flaky auth test" becomes
fix-the-flaky-auth-test, and follows renames. - No re-installs. Dependency folders (
node_modules,.venv,vendor,target) are shared into each chat's copy rather than downloaded again. - It's a real git worktree. You can
cdinto it and use plain git any time you like.
Then there are no branches and no Keep button — the agent just works in the folder, like any other program.
The browser
The chat sits next to a real browser, and Claude can drive it: open a page, click, type, read it back. Not a hidden browser it describes to you second-hand — the one on your screen, with your logged-in session. You watch it work, and you can take over at any moment.
- Your logins are your logins. Browser sessions live in a profile on your Mac, the same as any browser. That's what makes it useful on your real dashboard.
- Automation stays in here. The agent only ever drives Superagent's own browser pane, never your personal Chrome or Safari.
- Dev servers. Point it at localhost, or ask it to start the server itself — it stays one click away in the toolbar.
- One page, seen twice. One click shows the page as a desktop card and an iPhone frame at the same time, same session, both live.
The iPhone in the window
An iOS Simulator streamed inside the app from its own framebuffer — the real device, live at the rate it renders. Tap, swipe and type on it. So can the agent, which means it builds, launches and checks its own app without you playing courier.
- Nothing to install beyond Xcode: the input helper ships with Superagent.
- Apple's own Simulator window stays shut unless you ask for it, so you're never watching the wrong device.
- The pane reveals itself when the agent boots or launches something.
The simulator and the page you're on sit side by side on the same surface.
Files, @ and snips
- Click any file to read it — PDFs, images, markdown, source — right beside the tree. You can annotate PDFs in place, and drag files into the chat.
@reaches everything. Type@for this project's files, the other projects in your sidebar by name, or any folder on the disk (@/,@~/), drilling in one level at a time.- Snip to attach. The scissors on the browser or the simulator (or ⌘ ⇧ S): drag a box on the page or the phone screen and the crop lands in your message at full resolution.
- See every edit the moment it happens, with just the change highlighted.
What the agent is allowed to do
One picker in the composer decides how much rope it gets. It matches Claude Code's own modes, and you can change it mid-conversation.
| Mode | What it does |
|---|---|
| Full | Runs commands and edits files, like your terminal. The default. |
| Ask | Asks before commands and edits — approve on the Mac, or on your phone if that's where you are. |
| Edits | Applies file edits; some commands may be refused. |
| Plan | Read-only. It plans without changing anything. |
It's the mode that lets you learn what the agent wants to do before it does it, and it's the one that makes the phone useful.
Interject mid-turn
Type while the agent is working and it sees your message before it finishes — the same as the terminal. You don't have to wait for it to be wrong in peace.
The board, routines and the dashboard
A board the agent keeps
Each project has a list of work in four stages — todo, doing, testing, done — and Claude moves the cards as it goes. Watch them move while you talk to it. Add your own; it will pick them up.
Routines
"Check this site every hour," in plain language, on a timer. Routines run inside Superagent against this browser and your logged-in session, so they see what you'd see. They run while the app is open.
Dashboard
Turns per day, tasks done, a streak, and which projects actually got your time. Computed locally — the numbers never leave the machine.
Your phone
Pair the Superagent iOS app once from Settings → Phone and it follows your Mac from anywhere: live transcripts, send a prompt, start a chat, and approve what the agent asks — with a push notification when it's done or needs you.
How to pair
- Open Settings → Phone on the Mac.
- Scan the QR code with the iPhone app.
- That's it. The key in that QR is the only key that exists.
How private it actually is
Everything between the phone and the Mac is end-to-end encrypted with a per-device key from the pairing QR. Both sides dial out to a tiny blind relay that forwards ciphertext and stores nothing, which is why it works behind any network with no setup. You can run your own relay with one command and change the URL in Settings.
The Mac opens its relay connection when a phone is paired or being paired. An install that never pairs a phone never connects anywhere.
Note: the iOS app is built from source for now — see its README.
Shortcuts & small things
| ⌥ Space | Hold to talk. Transcribed on your Mac. |
|---|---|
| ⌘ ⇧ S | Snip the browser or the simulator into your message. |
- It never steals your focus. Agents finish quietly in the background; the window comes forward only when you ask.
- Notifications that say something — done, or has a question, with a summary of the agent's actual last reply.
- A context gauge on every conversation, showing how much of the context window it has used.
- Quiet by default. A burst of activity folds into one line you can open, instead of a wall of noise.
- Light on memory. Sessions start on your first message and wind down when idle; preview panes are released when a chat is in the background.
- Light and dark, following your system.
What leaves your Mac
The honest answer, in full. There is no Superagent server, so there is nowhere for your work to go but your own disk — and you don't have to take that on trust, because every line is public.
| Connection | When |
|---|---|
| Claude Code → Anthropic | Its own connection, on your subscription. Nothing of ours in between and no second bill. |
| Release check | This repo's releases, to see whether there's an update. |
| The companion relay | Only once you pair a phone. One outbound connection, ciphertext only. |
| The dictation model | Once, the first time you hold ⌥ Space. Whisper's weights come from Hugging Face and are cached on your Mac; the transcription itself never leaves it. |
Nothing else. No account, no sign-up screen, no analytics, no telemetry SDK in the build. Your browser logins live in a profile on your Mac, the same as any browser.
Read it yourself
All three repositories are MIT, every line: the desktop app, the iPhone app and the relay. Builds are signed and notarized by Apple, so the one you download is the one that was published.
When something's wrong
The agent won't answer at all
Superagent runs Claude Code; it isn't an AI of its own. Open Claude Code on its own and check it responds and is signed in. If it doesn't work there, it can't work here.
The simulator pane never appears
It needs Xcode installed, and it reveals itself when the agent boots or launches something — there's no button to summon an empty one. Ask the agent to boot a device.
The simulator is frozen on one picture
A shut-down device keeps its last frame. Ask it to boot again; the stream restarts with it.
A page in the browser pane isn't logged in
The pane has its own browser profile on your Mac, separate from Chrome or Safari. Sign in once inside the pane and it stays signed in.
My phone won't pair
Pairing needs both devices online at the same time. Re-open Settings → Phone to generate a fresh QR — the old one's key is gone once you leave the screen.
A chat's changes have vanished
Each chat works on its own branch, so its changes aren't in your main checkout until you press Keep. If you pressed Throw away, they're gone on purpose.
Something else
Bugs, roadmap and arguments all happen in public — open an issue.
Questions
Do I need to know how to code?
No. That's the point of the browser pane and the simulator: you judge the work by looking at it. The code is there when you want it, and ignorable when you don't.
Do I need an API key?
No. It runs on the Claude subscription you already have. There's no key to paste and no second bill.
Is it a Claude Code wrapper?
It's Claude Code with a computer: a browser on your logins, an iPhone, a file tree, a board, a scheduler and a phone in your pocket. The agent can finally do things rather than print them.
Windows or Linux?
Not today — it's a Mac app, Apple Silicon, and the simulator is a macOS thing. If cross-platform matters more to you than the browser and the phone, use Orca.
Can it use agents other than Claude Code?
Claude Code today. Codex and Antigravity are next — the plan is an agent layer other CLIs and local models can plug into.
Will it touch my personal browser?
Never. Automation only happens in Superagent's own browser pane.
Is my code sent anywhere?
Only where Claude Code itself sends it, on your own subscription. There is no Superagent server. See what leaves your Mac.
What does it cost?
Nothing. It's free and MIT-licensed. You pay Anthropic for Claude, as you already do.
Something here wrong or out of date?
Tell us
— these docs live in the repo, under docs/docs/.
MIT licensed · superagent.computer