Connect other AIs

Let Claude, ChatGPT, Cursor, Codex or GBrain read your to-dos and Graph over MCP, and with Read & edit make changes you can undo.

1. What another AI can do
2. Connect one
3. Official app or private address
4. What it sees
5. Messages
6. In Settings
7. Keep it safe

What another AI can do

Other AIs can connect to your Machine over MCP: Claude, ChatGPT, Cursor, Codex, GBrain and others. They can read your to-dos, promises, team briefs and the people and companies in your Graph.

With Read & edit they can also add a to-do, or a to-do with an email draft, and change your Graph: save a fact about a person or company, fix their details, fields and tags, or add a line to a person's summary.

Each change shows in Activity with Undo. They never merge, delete or share records. What they save shows AI: and their name, in their area. You can edit or remove it, and The Machine may revise a summary line later. Your Machine treats it as their claim, never as your instruction.

A draft waits in your Machine until you read it and press Send yourself. A connected AI never sends anything and never acts for you. Your Machine's own Do it and agents are not part of this.

Connect one

Open Settings, AIs. There are two ways in. Official app comes first: one address for every Machine, https://mcp.themachine.computer/mcp, with Copy. Below it is your Private address (end-to-end), https://<your name>.mcp.themachine.computer/mcp.

On a Machine on your own computer, press Get your address first. A page on your account opens: check that it shows the same code, confirm and pick a name. Your Machine pairs by itself, no restart. Both ways need this.

It then gets its own certificate for the address, which takes a few minutes and needs the cryptography package for Python.

In ChatGPT, open chatgpt.com/plugins, press +, then Add custom MCP server. Name it The Machine, paste the address and pick OAuth. Confirm the warning and press Create as a plugin. Your plan or workspace has to allow custom MCP servers.

In Claude, open Customize, Connectors, press + Add, then Add custom connector. Name it The Machine, paste the address, press Continue, then Add. On Team and Enterprise, an owner adds it in Organization settings.

With the official app, the AI asks you to sign in with your The Machine account. Pick this Machine if you have more than one, then approve it on your Machine's own page. With the private address, the AI comes straight to your Machine.

AIs on a device in your tailnet, like Claude Code or Cursor, can also use your dashboard's address followed by /mcp. Unpair takes a Machine on your own computer off your account again.

Your Machine shows which AI it is. An AI it can't recognize shows the address it came from: connect only if you just started this yourself.

Pick what it can do, Read or Read & edit. With both Work and Personal accounts, pick what it can see. Then pick for how long: 1 day, 7 days, 30 days or 1 year. Press Connect, or Cancel to send it back with nothing.

Official app or private address

The official app is the easy way: one address for everyone, and your account finds your Machine. The cloud routes its requests on the way but stores nothing.

With your private address, everything is encrypted all the way to your Machine. The cloud only passes it on and can't read it.

Either way your Machine decides. Only an AI you approved on its page gets in, with the access you picked. In Settings, AIs, an AI that came through the official app shows Official app.

What it sees

A connection sees only its area. Work covers your work accounts' to-dos and promises, work people, companies, projects and your Work Fleet's team brief. Personal covers the rest. Message text stays on your Machine unless you turn on Messages: without it, an AI sees a to-do's sources by sender, subject and date only.

Teammates' Team Graph records never go to an outside AI.

Messages

Messages lets an AI search and read your email, Slack, WhatsApp and other chats. It is off until you turn it on, on the page where you connect it or later in Settings, AIs. It only reads: it never sends, deletes or marks anything.

It reads only messages in its area. Some never go out: one-time codes and sign-in links, messages of an Only you board or project, threads The Machine judged sensitive (deals, HR, legal, health, money), threads you kept private, and deleted messages.

Every read shows in its Activity with how many messages and which threads, never the text. Your Machine hands the text over marked as data, so the AI shouldn't follow instructions written in a message.

While Messages is on, a Read & edit AI's Graph changes wait for your OK. Press Apply or Dismiss on the change in Activity.

In Settings

Open Settings, AIs to see each AI that connected, with when it was last used and how many calls it made. Change its access, its area or how long it lasts there. The next call uses it. Turn off now ends every sign-in it has at once. Activity lists every call, refused ones too. Remove takes it away.

An AI that can't sign in by itself can use a key. Under Add a key, name it, pick the access and how long, and press Make key. Copy the key right away: it shows only once. The AI sends it as Authorization: Bearer <key>.

Keep it safe

Every call is logged, refused ones too. A connection that expired or is off is refused until you change it. Signing in uses OAuth with PKCE. Your Machine keeps only a fingerprint of each token, so a copy of its data opens nothing.