Connect Claude or ChatGPT (MCP connector)
You can connect Claude or ChatGPT to your LoopCodeLab workspace. LoopCodeLab is then a connector in your AI app: a remote MCP server (Model Context Protocol) that you add once with one address. After that you work in plain words. You can talk an idea through, look at the plan, start the build, and ask how it is going.
At a glance
Section titled “At a glance”| Question | Answer |
|---|---|
| What is the address? | https://loopcodelab.com/mcp. You also find it in Settings under Connected apps |
| What can it do? | Ask the questions LoopCodeLab would ask about your idea, draw a mockup, show you the plan, start the build and report its progress. It can also see your builds, schedule, revise, pause and resume them, repair a failed build, manage stories and agents, check your setup, and show how much of your agent limits is used |
| Does it build straight away? | No. It talks the idea through first and starts only when you agree to the plan. Say “just build it” to skip that |
| Does it tell me about progress by itself? | Only while you are in the conversation. The assistant checks on the build and reports what changed. LoopCodeLab cannot send a message into your AI app when you are away. Ask “how is my build going?” at any time |
| What can it not do? | Open a terminal, delete a build that has started, or read or type a key. It can start a provider sign in that you then finish yourself |
| Can it delete anything? | Yes. It can cancel a scheduled build. It can skip a story, which deletes the work on that story. When it edits or retries a story, gives a story to another agent, or repairs a failed build, the attempt on that story is thrown away and the story is built again |
| Whose builds does it see? | The builds in your workspace, the same ones you see in the app. Never those of another workspace |
| Does it cost anything? | Starting or changing a build uses your credits or included builds, the same as in the app. Questions, plans and progress checks use none. Each mockup counts against your mockup allowance |
| Can I disconnect? | Yes, at any time, in Settings |
Connect
Section titled “Connect”- In LoopCodeLab, open Settings, then the Workspace & more tab. Under Connected apps, copy the address.
- In Claude, open Settings, then Connectors, and choose Add custom connector. Paste the address. In ChatGPT, add a connector the same way.
- Press Connect. A LoopCodeLab page opens. Sign in if asked.
- Read what the app asks for, then press Allow.
You are sent back to your AI app, and the connection is ready. Your AI app asks you to approve each tool the first time it is used.
If you do not see Connected apps in Settings, the connector is not open for your account yet.
From idea to build
Section titled “From idea to build”The assistant does not build the moment you mention an idea. It follows the same steps as the LoopCodeLab chat.
| Step | What happens | Uses credits? |
|---|---|---|
| 1. Questions | The assistant asks what LoopCodeLab would ask: who it is for, how it should look, whether people sign in, what content it needs | No |
| 2. Mockup | If you want one, it draws a picture of the look and shows it in the chat. You can ask for changes. It asks before each picture | No credits. Counts against your mockup allowance |
| 3. Plan | It shows the stories that would be built. Ask for changes until the plan is right | No |
| 4. Start | When you agree, it starts the build from that exact plan, with the mockup you chose | Yes |
| 5. Progress | It tells you when stories finish, when something needs attention, and gives you the preview link at the end | No |
To skip steps 1 to 3, say so: “Just build a one page site for my bakery.”
A plan you have seen is kept for one hour. After that, ask for the plan again.
What to ask
Section titled “What to ask”- “I want an app for my hair salon. Help me plan it.”
- “Show me a mockup of the home page.”
- “Change the plan: add online payment.”
- “Start the build.”
- “How is my build going?”
- “Why did the payment story fail?”
- “Pause the shop build.”
- “Build a blog for my bakery, and start it tomorrow morning.”
- “Is everything ready to build?”
- “Which of my agents are available?”
Follow a build
Section titled “Follow a build”Building takes several minutes. While you stay in the conversation, the assistant keeps checking and tells you each change. If you leave and come back, ask how it is going.
A build is not listed under its name until planning has finished. If you started one a moment ago and the assistant cannot find it, wait a minute and ask for your list of builds. Do not start it a second time.
When a name is already taken, LoopCodeLab adds a number to the new one: shop becomes shop-2.
Check your setup and your agents
Section titled “Check your setup and your agents”| Ask | You get |
|---|---|
| “Is everything ready to build?” | Which agents can lead a build, whether your included builds are usable, and any problem with a sign in or key, each with the fix and a link |
| “Which of my agents are available?” | For ChatGPT (Codex) and Claude: how much of the 5 hour and weekly limit is used, when each resets, and whether the agent is out. These are the provider’s own figures |
Limits are read from a subscription sign in, not from an API key. For Claude they need a full sign in: open a terminal in your workspace and run claude auth login. Other agents do not report their limits, so the assistant lists them as connected and not measured.
You see the same figures in Settings in the Agent limits card.
Sign in to a provider from the chat
Section titled “Sign in to a provider from the chat”When a ChatGPT, Grok or Kimi sign in has expired, you can renew it without opening LoopCodeLab.
- Say “sign me in to ChatGPT”.
- The assistant shows a link and a one time code.
- Open the link, sign in on the provider’s own page, and enter the code there.
- Come back and say “done”.
The assistant never sees your password or a key. Sign ins that need a code pasted back, such as Claude, are done in Settings instead.
This needs the permission Start provider sign ins. A connection made before that permission existed does not have it: remove the connection in Settings and connect again.
Permissions
Section titled “Permissions”When you connect, the LoopCodeLab page lists what the app asks for. You allow all of it, or you cancel. You cannot allow only a part.
| Permission | What it allows |
|---|---|
| See builds | Read your builds, their stories, status, logbook and preview links, the builds that are scheduled, and how many included builds you have left. Follow a build’s progress. Check whether you are ready to build, and see how much of the ChatGPT (Codex) and Claude limits is used. It never sees a key itself |
| Start and steer builds | Ask planning questions, draw mockups and plan a build for you to review. Start a build, schedule one or cancel a scheduled one, revise, pause, resume, repair a failed build, change the agent, and add, edit, retry or skip stories |
| Start provider sign ins | Start a sign in to ChatGPT, Grok or Kimi for your workspace and show you the link and the one time code. You finish it yourself on the provider’s page |
Your AI app may ask you to confirm before it does something that uses credits. We recommend leaving that confirmation on.
Disconnect
Section titled “Disconnect”Open Settings, then Workspace & more. Under Connected apps, press Remove next to the app. It loses access at once.
If something does not work
Section titled “If something does not work”When LoopCodeLab refuses something, the assistant tells you what is wrong, who can fix it, and the steps, with a link to the page where you fix it.
| What you see | What to do |
|---|---|
| “Not available yet” | Connecting AI apps is not open for your account yet |
| “This link has expired” | Go back to your AI app and connect again. A sign in link works for ten minutes |
| “We could not confirm it was you” | Go back to your AI app and connect again. Do not sign out or switch accounts in another tab before you press Allow |
| The assistant says a build could not start | Follow the steps and the link in its message, then ask again. You can also ask it to check your setup |
| The assistant says a sign in has expired | That is a model sign in inside LoopCodeLab, for example ChatGPT. It is not the connection to your AI app, so you do not need to connect again. Renew it from the chat, or follow the link in the message |
| The assistant cannot find a build you just started | Planning is still running. Wait a minute and ask for your builds. Do not start it again |
| The assistant says the plan is no longer available | A plan is kept for one hour. Ask for the plan again, then start |
| The mockup does not appear in the chat | It is still saved. Open Assets in LoopCodeLab to see it |
| The assistant cannot offer a provider sign in | Your connection was made before that permission existed. Remove it in Settings and connect again |
| The assistant stopped working after some weeks | Connect again. A connection that is not used for 30 days ends by itself |
| The assistant suddenly asks you to connect again | Connect again. LoopCodeLab closes a connection when it sees signs that someone else may be using it |
For developers
Section titled “For developers”| Item | Value |
|---|---|
| Protocol | Model Context Protocol, Streamable HTTP, one request per call |
| Endpoint | https://loopcodelab.com/mcp |
| Sign in | OAuth 2.1 with PKCE and dynamic client registration. Discovery at /.well-known/oauth-authorization-server |
| Scopes | builds:read, builds:write, account:manage |
| Tokens | Access tokens last one hour, refresh tokens 30 days and change on every use |
| Tools | 26. Each one calls the same LoopCodeLab API the web app uses, as the signed in user |