First, the words
Ten words you will see in this guide.
Read these once. Every step below uses them.
The same company. MindStudio was the drag-and-drop tool. Remy is its newer product, where an app is a folder of files instead of a canvas. You sign in at app.goremy.ai.
What Remy calls one project. An app can be a chatbot, a research agent, a form that does something clever, or a set of tools. Your "agent" is an app.
Anthropic's coding assistant. You talk to it in plain English and it writes the files, runs the commands, and fixes mistakes. It comes two ways: a desktop app with buttons, or a command you type in a terminal. This guide covers both.
A window where you type commands instead of clicking. On Windows it is called PowerShell. On a Mac it is called Terminal. The Claude Code desktop app has one built in (menu Views → Terminal).
Your app lives in one folder on your computer. "Repo" means that folder is tracked by Git, so every change is saved with a note and can be sent to Remy.
A free program that tracks changes to files and sends them to a server. You will not type Git commands. Claude Code does. You only need it installed.
A free program that runs JavaScript on your computer. Remy apps are written in TypeScript, a flavour of JavaScript, so Node has to be installed.
A small program called mindstudio-local. While it runs, the app on your computer behaves the way it would on the internet, with the same sign-in, database, and AI models. It gives you a preview link.
A long password that proves to Remy it is you. You will never need to see it. Signing in through the browser saves it in a hidden file, and the tools read it from there.
Sending your folder to Remy. About 25 seconds later your app is live at a public web address.
The plan
Two parts. Do Part A once. Do Part B for every agent you build.
Part A sets up your computer and ships the tiny Hello app that Remy gives every new project. Once you have seen that loop work, Part B is the same loop with your own idea.
- Install Git and Node.js
- Install Claude Code and open a folder
- Sign in to MindStudio from the terminal
- Make the Hello app in Remy (in the browser)
- Let Claude Code download it
- Start the tunnel and see the preview
- Change one line and publish it
- Tell Claude Code what the agent should do
- Build, test, and fix, in plain English
- Publish it, give it an address, and plug it back into Claude Code as tools
One limit before you start. The old MindStudio drag-and-drop agents cannot be edited from Claude Code. Nothing outside the web editor can change them. What Claude Code can build is the new kind of app, which is what Remy makes. And creating a brand-new app takes one click in the browser, because Remy needs you signed in for that. That click is Step 4.
Money: Remy is a paid plan, and every time an app calls an AI model it costs a small amount at the model's normal price. The Hello app costs well under a cent per greeting. Check your account has a balance before Step 6.
Quick path
The whole loop, once you have done it once.
This is the fast version for someone who has already been through Part A. The first time, follow the full steps below; they explain every click. After that, the whole thing is just these five lines.
- In app.goremy.ai, click New App, copy the app id from the address bar, and close the tab. Do not type in Remy's chat.
- Tell Claude Code: download my Remy app (that id) into a folder, and set its name in mindstudio.json. It handles the Git sign-in from the key that Step 3 saved.
- Run
mindstudio-localin the app folder. The first run shows a menu, not your app, so press q and run it again. Ask Claude Code for thehttp://127.0.0.1preview address and open it. - Tell Claude Code what to build or change, in plain English. Watch the preview; the tunnel picks up each save at once.
- Say "publish this app." Claude Code commits, pushes, waits for the release, and gives you the public address. Live in about fifteen seconds.
That is the entire method. Everything else in this guide is either the first-time detail behind one of those lines, or an extra you can add later: a readable web address, a list of who may sign in, or the app's tools inside Claude Code.
Part A · Set up once
From nothing to a live Hello app.
Install Git and Node.js.
Two free programs. Accept every default in the installers.
- Go to git-scm.com/downloads/win, download, run it, click Next until it finishes.
- Go to nodejs.org, download the version marked LTS, run it, click Next until it finishes.
- Restart your computer. This makes the new programs visible to every window.
- Go to nodejs.org, download the version marked LTS, open it, click Continue until it finishes.
- Open Terminal (press ⌘ Space, type Terminal, press Enter). Type
git --versionand press Enter. If a window offers to install command line tools, click Install. That installs Git.
git --version and press Enter, then node --version and press Enter. Each prints a version number like 2.47.0 or v22.11.0. If either says "not recognized" or "not found", the install did not finish. Run it again.Install Claude Code and open it in a new, empty folder.
Pick one of the two ways. The desktop app is easier if you have never used a terminal. Both do exactly the same things from here on.
- Download it: Windows or Mac. Install and open it. Sign in with your Claude account.
- Make a new empty folder somewhere easy to find, for example Documents → remy-apps.
- In Claude Code, click + New session. Above the message box, click Project folder and choose the remy-apps folder.
- You will type your messages in that box and press Enter to send them.
- Open a terminal. Windows: press the Windows key, type PowerShell, Enter. Mac: ⌘ Space, type Terminal, Enter.
- Install Claude Code. Windows: paste
irm https://claude.ai/install.ps1 | iexand press Enter. Mac: pastecurl -fsSL https://claude.ai/install.sh | bashand press Enter. - Make the folder and go into it. Paste
mkdir remy-apps, Enter, thencd remy-apps, Enter. - Type
claudeand press Enter. Sign in when it asks. You now type messages at the prompt and press Enter to send.
Now send Claude Code its first job. When it asks permission to run a command, say yes. It will ask often at first; that is normal.
I am new to this. Please check that git and node are installed and tell me the versions in one line each. Then install the two MindStudio tools with this exact command: npm install -g @mindstudio-ai/agent @mindstudio-ai/local-model-tunnel When it finishes, run "mindstudio --help" and tell me in one sentence whether it worked. Explain anything that went wrong in plain words.
mindstudio command works. If it says npm is missing, Node.js did not install; go back to Step 1. Windows note: if you ever run this install command again later, redo the Windows fix in Step 6, because reinstalling the tunnel undoes it.Sign in to MindStudio from the terminal.
This is the step that saves your API key for you. You do it yourself, in a terminal, because a browser window has to open and you have to click in it.
- Open the built-in terminal: menu Views → Terminal (or press Ctrl `; on a Mac ⌘ works too). A terminal pane appears, already inside your folder.
- Type
mindstudio loginand press Enter. - Your browser opens a MindStudio page headed Connect to MindStudio. Who will be using this connection? with your own name as the first choice. Sign in first if it asks (same account as Remy). Pick your own name (Connect as yourself) and click Continue.
- The browser page may then say "Connection failed". Ignore it; the terminal is the truth. Go back to Claude Code: the terminal says Authenticated successfully and names the file it saved the key in.
- Open a second terminal window (leave Claude Code running in the first). Windows: Windows key, PowerShell. Mac: Terminal.
- Type
mindstudio loginand press Enter. - Your browser opens a MindStudio page headed Who will be using this connection? Sign in first if it asks. Pick your own name (Connect as yourself) and click Continue.
- The browser page may then say "Connection failed". Ignore it. Back in the terminal it says Authenticated successfully. Keep this window; you will use it again in Step 6.
mindstudio whoami and press Enter. It prints your name and your organisation. Your key is now saved in a hidden file in your home folder (.mindstudio/config.json). Every tool in this guide reads it from there. You never have to copy it anywhere.Make the Hello app in Remy. This is the only thing you do in Remy's editor.
Remy gives every new app a tiny working program: you type a name, an AI writes you a greeting, and the greeting is saved. That starter is the "empty project". You are going to take it over with Claude Code.
- In your browser go to app.goremy.ai and sign in.
- Click + New App (top right). Remy does not ask for a name. It creates an app called "New App" and opens it. Claude Code will name it in Step 5.
- Remy opens a quick-start page with a chat box. Do not type anything in that chat. If you do, Remy's own builder starts working on a hidden copy of the app, and later it will fight with Claude Code over which version is real.
- Look at the address bar. It looks like
https://app.goremy.ai/apps/a1b2c3d4-e5f6-7890-abcd-ef1234567890/edit. The long code between/apps/and/editis the app id. Select just that code and copy it. - Paste it somewhere safe for a minute (a note, a blank document). Then close the Remy tab.
https://git.mscdn.ai/<that code>.git, and its public address will be https://<that code>.madewithremy.com once you publish.Let Claude Code download the app onto your computer.
Replace the red part with your app id. Everything else in this message tells Claude Code how Remy's Git server wants to be spoken to, so you do not have to know.
Download my Remy app into a new folder here called hello-agent. The app id is PASTE-YOUR-APP-ID-HERE.
The Git address is https://git.mscdn.ai/<app id>.git. It needs HTTP Basic authentication: username is the letter x, password is my MindStudio API key. The key was saved by "mindstudio login" in the file ~/.mindstudio/config.json (if the MINDSTUDIO_API_KEY environment variable is set, use that instead). Rules:
- Never show me the key, never put it in the Git address, and never write it into any file inside the project.
- Use a small askpass script (GIT_ASKPASS) that prints x for the username and the key for the password, set GIT_TERMINAL_PROMPT=0, and pass -c credential.helper= so no credential popup can hang the command. Put a 120 second timeout on the clone.
- After cloning, also run "git fetch origin _draft" but do not merge or touch that branch; it belongs to Remy's own builder.
Then open mindstudio.json and change "name" from "New App" to "hello-agent". Do not change anything else in that file. That name becomes the app's name on Remy the first time we publish.
When it is done, list the files and explain in plain English, in five sentences or fewer, what mindstudio.json, src/app.md, and dist/ are for.
mindstudio.json (the list of what the app contains), a folder src (the description of what the app should do, written in plain Markdown), and a folder dist (the code that does it). If Claude Code reports "401" or "authentication failed", the key was not found: repeat Step 3 and try again.hello-agent/ mindstudio.json the list: app id, name, tables, methods, screens src/app.md what the app does, in plain words (this is the boss; code follows it) src/interfaces/web.md what the screen should look like src/interfaces/@brand/ colours, fonts, tone of voice dist/methods/src/helloWorld.ts the one job: ask the AI for a greeting, save it dist/methods/src/tables/default.ts the one table: name + greeting dist/interfaces/web/ the web page people see
From now on, tell Claude Code to work inside this folder. In the desktop app, start a new session and pick hello-agent as the project folder. In the terminal, type cd hello-agent then claude.
Start the tunnel and look at your app.
The tunnel needs a real terminal window of its own, because it draws a little dashboard and stays running. Leave it open the whole time you work. The first time, you will start it twice: once to sign in, once to open your app. Earlier versions of this guide missed that, and the tunnel looked broken when it was not.
- Open Views → Terminal. It opens inside the current project folder. If you are still in remy-apps, type
cd hello-agentand press Enter. - Type
mindstudio-localand press Enter. - First time only: a logo appears with Press any key to Connect Account. Press a key. Your browser opens the same "Who will be using this connection?" page as Step 3; pick your own name and click Continue. The terminal changes to ● Connected to Cloud and shows a menu (Open Agents, AI Model Providers, Actions).
- That menu is not your app. Press q to quit, then type
mindstudio-localagain and press Enter. - Now it says Starting dev session, installs the app's parts (a minute or two), and shows the app dashboard.
- Use the second terminal window from Step 3. Type
cd remy-apps/hello-agentand press Enter. (On Windows, if PowerShell opened in a different place, typecd ~/remy-apps/hello-agent.) - Type
mindstudio-localand press Enter. - First time only: at Press any key to Connect Account, press a key. In the browser page, pick your own name and click Continue. The terminal says ● Connected to Cloud and shows a menu.
- That menu is not your app. Press q, then type
mindstudio-localagain and press Enter. - Wait for Starting dev session to finish and the app dashboard to appear.
MindStudio Local Tunnel v0.5.104 ● Connected to Cloud Config: ~/.mindstudio-local-tunnel/config.json Open Agents Agents you are editing in the MindStudio IDE will appear here. AI Model Providers Ollama Installed (not running) Actions ❯ Connect to Agent - Connect your local editor to a MindStudio interface or script Sync Models Exit Up/Down Navigate • Enter Select • q Quit
MindStudio Local Tunnel v0.5.104 ● Connected to Cloud hello-agent ● main /v2/<your app id>/run?dev-preview=true Session ID: df31fbb1-9f48-4fd3-b8b3-6adfa32327eb Release ID: df31fbb1-9f48-4fd3-b8b3-6adfa32327eb Branch: main App URL /v2/<your app id>/run?dev-preview=true localhost:5173 ● running greetings No requests yet Actions ❯ Sync Schema - Re-sync table definitions from disk Stop Session Local Models Up/Down Navigate • Enter Select • q Quit Screen Info Requests Methods Dev Server ←/→
Finding the preview address. The dashboard does not show the address you need. The App URL line points at a page MindStudio has retired, and localhost:5173 is the raw page with no connection to your account, so its button does nothing. The tunnel serves the real preview at http://127.0.0.1: followed by a port number that stays the same for your app but is never printed. Ask Claude Code to find it once and keep the link.
The tunnel (mindstudio-local) is running in another window for this app. Find the local address where it serves the preview: the port on 127.0.0.1 owned by the tunnel's own node process (the one running local-model-tunnel/dist/cli.js), not the 5173 dev-server port. Confirm by fetching it: the page must contain the text __MINDSTUDIO__. Give me the http://127.0.0.1:<port> link on its own line. Do not stop or restart anything.
mindstudio-local again. Macs skip this box. Reinstalling the tunnel undoes the fix, so paste it again after any reinstall.The MindStudio tunnel fails on Windows with "Only URLs with a scheme in: file, data, and node are supported ... Received protocol 'c:'". Find the tunnel's dev-worker.js under the global npm folder (npm root -g, then @mindstudio-ai/local-model-tunnel/dist). Back it up with today's date in the name. In the line that does import(transpiledPath + "?v=" + version), convert the path with pathToFileURL from node:url before importing. Do not touch anything else. Tell me exactly what you changed, then tell me to start the tunnel again.
http://127.0.0.1: address in your browser. A dark page says Hello World, AI-powered greetings, with a name box and a Say Hello button. Type a name and click. A greeting appears within about ten seconds. It was written by an AI model through your MindStudio account, and the code that asked for it ran on your computer. In the tunnel window, press the right arrow to the Requests tab: it lists ✓ helloWorld with a time in milliseconds. A ✖ helloWorld with the "Only URLs with a scheme" message means the Windows fix above has not been applied yet.Request Log✖ helloWorld 350msOnly URLs with a scheme in: file, data, and node are supported by the default ESM loader… (code: ERR_UNSUPPORTED_ESM_URL_SCHEME) Request Log ✓ helloWorld 5707ms
If you have used the tunnel before this year, read the first item in If it goes wrong before you press a key.
Change one thing, see it, then publish it.
This is the whole loop. Once you have done it, you can do everything else.
Make one small change to this app: the greeting should also mention the day of the week. Update src/app.md first so the description matches, then the method in dist/methods. Do not change anything else. Tell me when it is done and what file you changed.
Now send it to the internet.
Publish this app. Steps: 1. Commit every change with a short clear message. 2. Push to origin main using the same Git authentication as when you cloned (askpass script, no credential helper, 120 second timeout). 3. Wait for Remy to build it: run npx -y @madewithremy/admin releases wait --commit <the commit you just pushed> with two environment variables set for that command only: WORKSPACE_DIR = this folder's full path, and MINDSTUDIO_API_KEY = the apiKey value from ~/.mindstudio/config.json (the admin tool does not read that file by itself). Never show me the key. Then run npx -y @madewithremy/admin releases list the same way and tell me the status. 4. Give me the public address, which is https://<app id>.madewithremy.com, and confirm it answers by fetching it. Report three things separately: pushed yes/no, release status, public address checked yes/no.
Part B · Every new agent
Now build the one you actually want.
Do Step 4 and Step 5 again with a new name to get a fresh app (or keep working inside hello-agent). Then:
Give Claude Code the house rules, then describe the agent.
First, a short rules file. Claude Code reads a file called CLAUDE.md in the folder at the start of every session. Paste this once per app.
Create a file called CLAUDE.md in this folder with these rules, then follow them: - This is a MindStudio Apps v2 (Remy) app. The description in src/app.md is the boss; update it before changing code. - Before writing any MindStudio SDK call, run mindstudio ask "<your question>" in the terminal to get real model names and the exact way to call an action. Do not guess model names. - The tunnel (mindstudio-local) is running in another window. Test every change through its preview before telling me it works. - Every AI call costs money. Before anything that runs in a loop or over a list, run mindstudio estimate-cost <action> and tell me the estimate first. - Never show me API keys and never write them into project files. - Publish only when I say "publish". Publishing means: commit, push to origin main with the askpass method, wait for the release, and check the public address. - Explain things to me in plain words. I am not a programmer.
Then describe what you want, the way you would to a colleague: what goes in, what should come out, and where the results should be kept. Use the fill-in-the-blanks version below; the red parts are yours.
I want to build an agent that does what, in one sentence. Who uses it: just me / my team / the public. What the person gives it: for example a name and a company, a web link, a pasted document. What it should give back: for example a short report with sources, a list, an email draft. Where results should be kept: save them so I can see past results / do not save anything. Which outside things it needs: for example web search, reading a web page, Gmail, Slack, a Google Sheet, none. How it should be used: a web page / from Claude Code as a tool / on a schedule every morning / all of these. Before you build anything, ask me up to five questions about the parts that are unclear. Then write the plan into src/app.md and show it to me. Do not write code until I say the plan is right.
A real example, filled in the same way:
I want to build an agent that researches one person before I meet them. Who uses it: just me and one colleague, with sign-in by email code. What the person gives it: the person's name, plus a LinkedIn link or their company, and a line about why I am meeting them. What it should give back: a one-page dossier: their role and career, recent news with dates and links, three talking points for my meeting, and a list of sources. Where results should be kept: save every dossier so I can reopen it, and do not research the same person twice within 30 days. Which outside things it needs: web search and reading web pages. How it should be used: a web page, and also as a tool I can call from Claude Code. Use claude-5-sonnet, and if that model errors, fall back to claude-4-6-sonnet automatically. Show the cost of each run in dollars.
Build, test, and fix, in plain English.
Claude Code writes the files and tries them through the tunnel. You look at the preview and tell it what you see. The messages below do most of the work.
Run every method once through the tunnel with realistic test input. Show me each input, what came back, how long it took, and what it cost. Fix anything that fails and run it again.
Estimate what one full run of this agent costs in dollars, and what 100 runs would cost. Tell me which step is the expensive one and whether a cheaper model would do for it.
On the preview page I did this: what you clicked or typed. I expected what you expected. Instead I saw what actually happened, copied exactly. Find the cause in the tunnel's request log, explain it in one sentence, fix it, and tell me what to try again.
Change: what you want different. Update src/app.md first, then the code. Do not change anything I did not ask about. Tell me which files changed.Publish, give it a proper address, and plug it into Claude Code.
Publishing is the same message as Step 7. After it is live, three optional messages make it nicer to use.
Give this app the address my-agent.madewithremy.com using npx -y @madewithremy/admin domains set my-agent (WORKSPACE_DIR set to this folder on Windows). Then fetch the new address and confirm it answers.
Only these email addresses may sign in: you@example.com, colleague@example.com. Use npx -y @madewithremy/admin settings allowlist add <email> for each, then settings allowlist enable. Show me the final list.Add an MCP interface to this app so its methods can be used as tools from Claude Code. Create src/interfaces/mcp.md and dist/interfaces/mcp/interface.json, add the mcp entry to mindstudio.json, and expose only the methods a stranger would need, each with a description that says when to use it and when not to. Mark read-only tools as readOnly. Then publish. When it is live, add the server to my user-level Claude Code settings with claude mcp add --scope user --transport http <app name> https://<app address>/_/mcp --header "Authorization: Bearer <my MindStudio API key>" reading the key from ~/.mindstudio/config.json without showing it. On Windows run that command from PowerShell, not from Git Bash. Then tell me to restart Claude Code.
/mcp. Your app is listed as connected, with its tools. Now, in any session, you can say "research Jane Smith at Acme before my 3 pm meeting" and Claude Code calls your agent. The same app can also answer as a web page, on a schedule, from a web form, or by email; ask for those the same way.Already built
Apps and agents you already have.
Part A starts from an empty app. Most people arrive with things already built, in two very different generations. This section covers both kinds, with proof.
A Remy app you built in Remy's own builder.
Every Remy app, however it was made, is a Git repository on Remy's server, so Claude Code can take any of them over the same way as the Hello app. Nine apps in one workspace were checked this way on 4 September 2026; six of them had been built entirely in Remy's chat, months earlier, and every one downloaded with its spec, methods and screens intact.
- Go to app.goremy.ai. All Apps lists every Remy app in your workspace, with the date each was created and last deployed.
- Click the app. The address bar shows its id between
/apps/and/edit, exactly as in Step 4. Copy it, then close the tab. Do not type in Remy's chat while Claude Code works on the app. One author at a time. - Paste the message below into Claude Code, with the id and a folder name of your choosing.
Download my existing Remy app into a new folder here called FOLDER-NAME. The app id is PASTE-YOUR-APP-ID-HERE. The Git address is https://git.mscdn.ai/<app id>.git with HTTP Basic authentication: username x, password my MindStudio API key from ~/.mindstudio/config.json (or the MINDSTUDIO_API_KEY environment variable if set). Use an askpass script, GIT_TERMINAL_PROMPT=0, -c credential.helper= and a 120 second timeout, and never show me or store the key. This app was built in Remy's own builder, so after cloning run "git fetch origin _draft" and tell me, in plain words, whether _draft has work that main does not (count the commits each way). Do not merge or change anything yet. Then read mindstudio.json and src/app.md and tell me in five sentences what this app does, what screens it has, and which tables it keeps.
mindstudio.json, src, dist), only bigger, and a plain summary of what the app does. If Claude Code reports that _draft is ahead of main, Remy's builder has unpublished work: decide which version you want before editing (the second item in If it goes wrong). From here, Step 6 (preview through the tunnel) and Step 7 (change, then publish) work unchanged. On a live app, make one small change first and publish it, so you know the round trip works before you ask for anything big.
Downloaded, changed, published, and restored from Claude Code, signed in as an agent identity.
- What happened
- The app was downloaded with the key from a "Claude Code · Agent" sign-in. One line of its sign-in page was changed, pushed, and released in 28 seconds. The live page, fetched without signing in, showed the new footer. The change was then reverted and pushed again; the second release restored the original page, and the repository is byte-for-byte what it was before.
- Before
- Footer: "Signal Desk · Operator instance" (left).
- After
- Footer: "Signal Desk · Operator instance · edited from Claude Code 4 Sep 2026" (below).
An old drag-and-drop agent.
These live in the older editor at app.mindstudio.ai, and they are a different kind of thing: a canvas of blocks, not a folder of files. No tool outside that editor can change the blocks, the prompts, the models or the triggers. Three things are possible from Claude Code.
Run "mindstudio agents" and show me the names, ids and descriptions as a table. (On 4 September 2026 that listed 51 agents in one workspace.)Run my MindStudio agent called NAME with this input: …. Claude Code uses mindstudio run-agent. The agent must be set up for API use in its own settings, or the run is refused.- Open the agent in the MindStudio editor and leave that tab open.
- In a terminal, go to a folder that is not an app folder (for example
cd ~/remy-apps) and runmindstudio-local. Without amindstudio.jsonnearby, the tunnel shows its agent menu instead of an app. - The agent appears under Open Agents. Choose Connect to Agent, then the agent.
- The tunnel lists the agent's custom interfaces and scripts. Pick one: it downloads that piece into a folder and shows a Copy Claude Code Command button. Paste that command into Claude Code, edit, then press Compile in the editor.
- If it says No interfaces in this agent and No scripts in this agent, there is nothing in that agent Claude Code can edit. Everything else in it belongs to the editor.
Open Agents Research Blogger Pro AI Model Providers Ollama Installed (not running) Actions ❯ Connect to Agent - Connect your local editor to a MindStudio interface or script Sync Models Exit Choose an Agent Don't see your agent? Make sure it's open in the MindStudio editor. ❯ Research Blogger Pro - Connect to this Agent https://app.mindstudio.ai/agents/e434f2a0-…/edit Refresh Back
Research Blogger Pro https://app.mindstudio.ai/agents/e434f2a0-…/edit Select an interface or script to connect your local editor. Interfaces No interfaces in this agent. Scripts No scripts in this agent. ❯ Refresh Back
Cost
What it costs, and what you save.
There are two separate meters, building an app and running an app, and building with Claude Code is where the money is saved.
Running an app costs whatever AI models it calls, at each model's normal price, charged to your Remy balance. An app with no AI in it, a page that shows a fixed message, runs for nothing. An app that calls a model every time someone uses it costs a small amount each time, and your Usage and Cost page keeps the running total for each app.
Building is where the two ways part. When Remy's own chat builds for you, it runs several strong models to write the code, and that work is charged to your balance as well. When Claude Code builds, the writing is done by Claude on your Claude plan, so your Remy balance barely moves. The same app, built the two ways, leaves a very different bill.
In plain figures: a tiny app built by Claude Code adds about a cent to the Remy balance, most of that just the platform stamping out the release. The same tiny app talked into being through Remy's chat runs to a few dollars, and a larger app built that way to far more, because every model it thinks with is billed to you. Running costs are identical either way, since the app's own AI calls are the same whoever wrote the code. So the trade is simple: you keep everything Remy gives you, the hosting, the sign-in, the database, the one-push publishing, and you move the cost of writing the code off your inference balance and onto a plan you already pay for.
One habit protects the saving: close Remy's chat editor while Claude Code works on an app, because Remy's builder wakes up and builds in parallel if you leave it open, and that parallel work is billed.
FAQ
What each method can and cannot do.
Remy is in beta. It launched publicly in June 2026 and changes often: button names, the tunnel's screens, and the limits below can all change without notice. Everything on this page was checked on 4 September 2026 against tunnel 0.5.104 and SDK 0.1.117. When the screen in front of you disagrees with this page, trust the screen, and tell Claude Code exactly what you see.
| Method | Can | Cannot |
|---|---|---|
| Claude Code with the app's folder (this guide) | Build a new Remy app from an empty one. Take over any Remy app in your workspace, however it was built. Edit spec, methods, screens and tables. Preview through the tunnel. Publish with one push. Name the app, set its address, restrict sign-ins, add tools for Claude Code. | Create the app itself (that needs your browser, Step 4). Touch an old drag-and-drop agent's blocks. Work safely while Remy's chat edits the same app. |
| Remy's own builder (the chat at app.goremy.ai) | Create apps, build and change them by conversation, publish them. | Import an old drag-and-drop agent. Share an app with Claude Code at the same time: it keeps a hidden _draft copy that will fight your pushes. Pick one author per app. |
The tunnel (mindstudio-local) | Run a Remy app on your computer with a live preview, real sign-in, database and AI. Show the request log. Connect to an old agent's custom interface or script. Lend your local AI models to the cloud. | List your apps (it only sees the folder you start it in). Print the preview address (ask Claude Code for it, Step 6). Change anything in an old agent beyond its custom pieces. |
The mindstudio tools inside Claude Code (188 tools, from mindstudio mcp) | Run any old agent that allows API use. Run single actions: search, scrape, generate text or images, send messages, and 850 or so integrations. | Build, edit, publish or delete anything. See Remy apps at all. |
| Your app's own tools (the MCP interface from Step 9) | Expose that one app's methods as tools Claude Code can call in any session. | Reach any other app; each app publishes its own. |
The admin tool (remy-admin) | Watch releases, set the address, manage the sign-in allowlist, read request logs, run a method on the live app. | Read the key that login saved; it needs MINDSTUDIO_API_KEY in the environment. Create or delete apps. |
| Old drag-and-drop agents | Run from Claude Code. Hand a custom screen or script to Claude Code through the tunnel. | Have their blocks, prompts, models or triggers changed by anything but the editor. Be moved into Remy; they are rebuilt from their description. |
If it goes wrong
Ten things that happen, and what fixes each one.
Each of these happened during a real build. None of them means you broke anything.
Older tunnel settings can be missing one field, and the sign-in may overwrite your saved key. Before you press a key at the logo screen, make a copy of the settings file.
Make a backup copy of ~/.mindstudio-local-tunnel/config.json with today's date in the name. Do not show me its contents.
Remy's builder saves its work to a hidden branch called _draft every five minutes while its editor is open. Close the Remy tab and let Claude Code look.
Run git fetch origin _draft and show me, in plain words, how the _draft branch differs from main. Do not merge anything yet. Then recommend which version to keep.
A model name can be down for a day on MindStudio's side. On 1 September 2026 claude-5-sonnet failed all day and worked again the next morning.
Test the model this app uses with a one-line generateText call. If it fails, try claude-4-6-sonnet, gpt-5.5 and gemini-3.7-flash, tell me which ones work today, and add an automatic fallback so the app tries the next one when the first errors.
Windows' Git Credential Manager has popped up a login box that nobody can see. Press Ctrl C in the terminal to stop it.
The push hung. Retry it with the askpass script, GIT_TERMINAL_PROMPT=0, -c credential.helper= and a 120 second timeout, exactly as in the clone step.
npm and npx sometimes stall on Windows. The programs are fine; the launcher is slow. Never let anyone "kill all node processes" to fix it: that also kills Claude Code's own helpers.
That command is stuck. Stop only that one process by its id, then run the same tool directly with node from node_modules (for example node node_modules/typescript/bin/tsc) with a timeout. Do not kill every node process.
Three causes. You opened localhost:5173 or the App URL line instead of the http://127.0.0.1: address from Step 6. Or the tunnel window shows the menu (Open Agents, Connect to Agent) instead of your app: press q and run mindstudio-local again inside the app folder. Or the tunnel window closed: run it again there. It should say Connected to Cloud, name your app, and list requests.
Long jobs need two small guards: refuse to start the same job twice while it is running, and mark a job failed if it is silent for 35 minutes.
Add two guards to the long-running job: (1) if the same request is already in progress, return that one instead of starting another, and (2) when reading a job that has shown no progress for 35 minutes, mark it failed. Test both through the tunnel.
MindStudio reports task-agent cost in nanodollars. Divide by 1,000,000,000 to get dollars. A six-step research run came to $0.42.
Wherever this app stores or shows cost, convert usage.totalBillingCost from nanodollars to dollars (divide by 1e9) at the moment it is saved, and show two decimals.
Git Bash rewrites /c into C:/ inside the command. Run any claude mcp add command from PowerShell (Windows key, type PowerShell), not from Git Bash. The command itself is fine.
Tunnel version 0.5.104 has a Windows bug: when your code runs, it tries to load the file with a c: path and the loader refuses it ("Only URLs with a scheme in: file, data, and node are supported"). Found on 2 September 2026 and still there on 4 September; 0.5.104 is the newest version. Mac is not affected. The fix lives in Step 6, in the box marked Windows only. Reinstalling the tunnel undoes it, so paste it again after any reinstall. If you skipped that box, paste this now, then press q in the tunnel window and start mindstudio-local again.
The tunnel fails on Windows with "Only URLs with a scheme in: file, data, and node are supported ... Received protocol 'c:'". Find the tunnel's dev-worker.js under the global npm folder (npm root -g, then @mindstudio-ai/local-model-tunnel/dist). Back it up with today's date in the name. In the line that does import(transpiledPath + "?v=" + version), convert the path with pathToFileURL from node:url before importing. Do not touch anything else. Tell me exactly what you changed, then tell me to press q in the tunnel window and start mindstudio-local again.
Desktop app or terminal?
The same Claude Code, two front doors.
Everything in this guide works in both. The clicks differ in these places.
| You want to | Desktop app | Terminal |
|---|---|---|
| Start working in a folder | + New session, then Project folder above the message box, choose the folder. | cd into the folder, then type claude. |
| Type a command yourself (sign in, start the tunnel) | Menu Views → Terminal, or Ctrl `. It opens inside your project folder. | Open a second PowerShell or Terminal window so Claude Code keeps running in the first. |
| See the preview | Open the tunnel's preview address in the app's Browser pane or in your normal browser. | Open the address in your normal browser. |
| Check which tools are connected | Type /mcp in the message box. | Type /mcp at the prompt. |
| Give Claude Code permission to run commands | Click Allow when it asks. To ask less often, type /permissions. | Press y when it asks. /permissions works the same. |
| Install it | Windows · Mac. On Windows, Git for Windows is recommended (Step 1 covers it); without it Claude Code falls back to PowerShell. A Linux build also exists. | Windows: irm https://claude.ai/install.ps1 | iex · Mac: curl -fsSL https://claude.ai/install.sh | bash |
Both read the same CLAUDE.md rules file and the same tool settings, and both ask before running anything that changes your computer. If you start in one and switch to the other, nothing is lost.
If you already have old-style MindStudio agents
What Claude Code can and cannot do with them.
| You want to | Possible? | How |
|---|---|---|
| Run an old agent and get its answer | Yes | Ask Claude Code: "Run my MindStudio agent called name with this input" (it uses mindstudio run-agent). The agent must be set up for API use in its settings. |
| Edit the custom screen or the script block inside an old agent | Only those two pieces | Open the agent in the MindStudio editor. Run mindstudio-local in a folder that is not an app folder. Choose Connect to Agent, then the agent. It lists that agent's custom interfaces and scripts and offers a "Copy Claude Code Command" button. If it says "No interfaces in this agent" and "No scripts in this agent", there is nothing in it Claude Code can edit. Edit, then press Compile in the editor. See Already built. |
| Change an old agent's blocks, prompts or triggers | No | Only the web editor can. There is no command for it. |
| Move an old agent into Remy | Not automatically | Copy the old agent's description (or its exported JSON) and paste it into your Step 8 brief. Claude Code rebuilds it the new way. |
Evidence
Four builds, each checked from a logged-out browser.
An empty Remy app, downloaded, changed and published from Claude Code without touching the editor again.
- Question
- Could a Remy app be built from outside Remy's own builder at all, and would its Git server accept the saved key?
- What happened
- Downloaded the starter, replaced it with one method and one screen, tested through the tunnel, published.
- Result
- Live 24 seconds after the push. The public address showed the new screen in a private window. Remy's hidden
_draftcopy was found the same evening and reconciled, which is where the "do not type in Remy's chat" rule comes from.
This guide itself, followed step by step on a Windows PC, with no key in the environment and the tunnel's settings wiped first.
- Question
- Does Part A work for a reader who starts from nothing, or only for the person who wrote it?
- What happened
- The first pass found four things the guide had wrong: Remy asks for no app name; the tunnel's first run lands on a menu, not your app, until you quit and start it again; the dashboard never prints the preview address it actually serves; and the publish tool refuses to read the key that login saved. With each fix in place, the step was run again.
- Result
- Greeting through the tunnel in 5.7 seconds, the day-of-week change picked up without a restart, release live 21 seconds after the push, and the public address answered "Happy Friday" in a logged-out fetch. The app was named from its manifest and Remy's app list showed the new name.
The example brief from Step 8, built and shipped: web page, sign-in, saved dossiers, and five tools inside Claude Code.
- Question
- Could a real multi-step research agent be built this way, and survive a model failing on launch night?
- What happened
- Email-code sign-in for two people, a dossiers table, background research with web search and page reading, checked claims, a printable report, a PDF export, and an automatic model fallback. A review pass before publishing found 28 problems; most were real and were fixed with messages like the ones in Step 9.
- Result
- Live at research-people.madewithremy.com. First real run: $0.42, 12 sources, six tailored talking points and three common-ground items. Its tools now appear in Claude Code's
/mcplist.
A small greeting app, written by Claude Code and published to Remy in a single push.
- Question
- Does this hold up for an app someone actually keeps, not just the Hello demo?
- What happened
- Claude Code wrote the app, a web page plus one backend method that returns the greeting with no AI in it, then committed, pushed, and waited for the release. No back-and-forth was needed.
- Result
- Live about fifteen seconds after the push, at its own web address, with the greeting served by the app's own backend in under half a second. Build cost on Remy: about a cent. Running cost: nothing, because the app calls no model.
Evidence note: how the platform behaves is taken from MindStudio's own public code and documentation (the mindstudio-ai/remy and mindstudio-local-model-tunnel repositories) and from Claude Code's documentation for the desktop app. Costs, timings and failures are from the four builds above, on 1, 2 and 4 September 2026. Button names in Remy's editor may change; the address-bar method in Step 4 does not depend on them.