Skip to main content
Paige runs an MCP server — a standard way for an AI agent to use a product on your behalf. Connect one and it can do the things you’d otherwise do by hand: send a WhatsApp message, read your conversations, query and update your project’s tables, edit your bot’s code, build a flow, deploy, or draft a broadcast. It only ever does what you approved. It works with any MCP-compatible agent, because MCP is an open standard they all speak — Claude, ChatGPT, the OpenAI Codex tool, and Cursor are all supported. Everything here lives under Settings → API.
For the full list of what a connected agent can actually do — all twenty-six tools, the permission each one needs, and the guardrail that stops it sending a broadcast — see What a connected agent can do.
The API section is for owners and admins. If you’re a workspace member without admin access, it won’t appear in your Settings sidebar at all.
Prefer raw HTTP? You don’t need an agent at all — the same actions are available as a plain REST API. See the API reference.

​
Two ways to connect

Sign in — no key

For agents that support it, like Claude and ChatGPT. You click Connect in the agent, sign in to Paige, tick the projects and permissions, and approve. Nothing to copy or paste.

Paste an API key

For any client. Create a key under API Keys, then open the Connect an AI agent (MCP) card and pick your client. Three of the five tabs give you a snippet to copy; the other two give you the server address to paste into the client’s own connector settings.

​
What each client tab gives you

The Connect an AI agent (MCP) card has five tabs. They don’t all work the same way, because the clients themselves don’t: So Claude Code, Claude Desktop and Cursor are copy-paste. ChatGPT and Codex CLI are server-URL-only: you copy the address and configure the client yourself.
On the three copy-paste tabs, the snippet comes prefilled with your new key while the key is still on screen — copy the whole thing then. The secret is never shown again, so afterwards the snippet falls back to a YOUR_PAIGE_API_KEY placeholder you fill in yourself.
If Paige’s MCP server hasn’t been switched on for your environment yet, the card says so and the snippets show a <your Paige MCP server URL> placeholder rather than a real address.

​
Approving the connection

When you use the sign-in route, Paige shows you a consent screen before anything is granted.
1

Check the web address first

Before you look at anything else, check your browser’s address bar. It must be Paige’s own site at https://paigeme.dev. See Make sure it’s really Paige below.
2

Tick the projects the agent may use

You can pick several projects at once — you don’t have to approve the same agent again for each one. Your own projects are grouped under Personal, and each workspace you help run gets its own group headed by the workspace name. If you only have one project, it’s picked for you.
3

Narrow the permissions per project

Every project you tick starts with all the permissions the agent asked for. Open a project’s row to untick anything you don’t want it doing there — so the same agent might read conversations in one project and edit code in another.
4

Pick a default project

When you tick more than one, choose which is the default. That’s the project the agent uses when it doesn’t say which one it means.
5

Approve

The agent is connected, with exactly the permissions you granted, project by project.
A paused project can’t be ticked — the row says so. An agent can’t do anything in a paused project, so resume it first.
If any project you picked isn’t yours to connect — a workspace project where you’re only a regular member, for instance — the whole approval is refused. Nothing is granted, not even the projects that would have been fine. Untick the one you can’t connect and try again, or ask an owner or admin to do it.

​
Make sure it’s really Paige

Check the address bar, not the logo. Before you approve access, look at the web address in your browser: it must be https://paigeme.dev. The address bar is the one thing a look-alike page cannot fake, so it’s the only thing you should trust.The consent card also carries the Green Pages logo and the words Paige by Green Pages, which helps you recognise the page at a glance — but treat that as a familiar face, not as proof. Anyone can copy a logo onto their own page.If the address looks wrong, or you didn’t just ask an agent to connect, close the page and don’t approve.
The screen looks and works the same on a phone. One thing that isn’t a warning sign: the small icon your agent shows next to Paige in its own list of connectors is chosen by that agent, not by Paige, so it may look plain there for a while.

​
One connection, many projects

You don’t need a separate connection per project. A single connected agent can cover several projects, and each project has its own permissions — a project can’t borrow a permission from another project on the same connection. The headline part is what happens afterwards:
You can add or remove projects later without reconnecting the agent. Go to Settings → API → Connected agents, find the agent, and click Add project. It takes effect straight away — the agent is not re-approved, nothing is re-pasted, and the connection it already has simply starts covering the new project too.
From the same card you can:
  • Change a project’s permissions on that connection at any time.
  • Remove a project from the connection. It stops working on the agent’s very next action, while its other projects carry on untouched.
  • Disconnect the agent entirely. Its access stops immediately and it would have to connect and be approved again.
You can’t remove a connection’s last project — use Disconnect instead. Otherwise you’d be left with an agent that’s still connected but can’t be used or managed. You also can’t add a project that’s paused or deleted.

​
What the Connected agents card shows

For each connected agent you’ll see every project attached to it, and for each of those projects: its own permissions, its own last-used time, and whether it’s the default. If a connection also covers projects you don’t have authority over, they aren’t named to you — you just see how many there are, so you still know the connection’s full reach.

​
Workspace projects need an owner or admin

You can connect an agent to a workspace project, not just your personal ones — but only if you’re an owner or admin of that workspace.
This changed. Connecting an agent to a workspace project used to work for any member of the workspace. It now requires owner or admin. An agent can be granted powerful permissions — editing code, deploying, sending broadcasts — so connecting one to a shared project is left to the people who run it. If you’re a regular member, ask an owner or admin.
The same rule applies to the Connected agents card itself: on a workspace project, only owners and admins can see the connected agents or disconnect them. Being an admin of one project never lets you attach a different project you don’t control. On your own personal projects nothing changes — you manage your agents yourself. See Workspaces for roles and what each one can do.

​
How the agent picks which project to work on

There’s nothing to configure for this. The agent can ask Paige which projects your connection covers, and it gets back each project’s name along with exactly what it’s allowed to do there. From then on it names the project it means with each action.
  • Anything that changes something — sending a message, editing code, deploying, setting up a broadcast — must name the project explicitly once more than one is attached. A change can’t land in the wrong project by mistake.
  • Read-only things, like listing conversations, fall back to your default project when the agent doesn’t say.
Say the project name in your request — “deploy the Acme Support bot” — and there’s nothing to guess. A good agent will ask you rather than pick for you. And if you’ve just added a project and the agent still says it can’t reach it, ask it to look up your projects again; it doesn’t need reconnecting.

​
Connecting specific agents

Copy the one-line claude mcp add … command from the Claude Code tab and run it in your terminal. The Paige tools appear the next time you launch Claude Code.
Either route works. Signing in is the easy one: add the Paige server, sign in, pick your projects and permissions, approve. Or copy the config block from the Claude Desktop tab into your claude_desktop_config.json and restart the app.
Use the API key route. Create a key with the scopes you need, then copy the config block from the Cursor tab into ~/.cursor/mcp.json (or .cursor/mcp.json in your project) and reload Cursor.
Add a custom connector in ChatGPT and give it Paige’s MCP server address, shown on the ChatGPT tab of the card. ChatGPT connectors always sign in — its form only offers OAuth, No Authentication, or Mixed Authentication, so there’s no box to paste an API key — and it sends you to Paige to choose your projects and approve.
Copy the server address from the Codex CLI tab and add the Paige server to Codex’s own MCP config (its mcp_servers block). Codex can then either sign in to Paige and approve, or use a Paige API key you paste in — whichever you prefer.

​
When something doesn’t work