Connect your AI assistant

Once connected, an AI assistant you run elsewhere — Claude Code or Codex in your own terminal, or any MCP client — can see your machines and sessions, read and write shared memory, do work on other machines once you approve, and pause to ask for your review.

Sessions opened inside Kite already have these abilities. This guide is for assistants running outside Kite.

The quickest way: let it connect itself

Send this sentence to your assistant:

Read https://kite.lumenlan.com/skill.md and follow it to connect Kite

Three things happen next:

It asks for access and gives you a link and an 8-character pairing code, such as K7QM-4TXR.
You open the link (signed in to Kite), check that the code matches exactly, untick anything you do not want to grant, and choose Allow.
The credential goes straight to the assistant. It installs it and lists your machines to confirm.

You never handle a key. The pairing code is valid for 10 minutes; if it expires, ask the assistant to start again.

Only approve a request you just started yourself. If you were not expecting it, or the codes do not match, choose Deny.

What it can do

It can only do what you left ticked on the approval page:

Permission	What it can do
See your machines, their status and session content	List machines and sessions and read what happened in them
Message sessions, answer their questions, start new sessions	Before starting a session it tells you how much of your allowance it uses and waits for your confirmation
Read shared memory	Your name, habits, and which machine and folder each project lives in
Write and update memory	Everything it writes is visible on the Memory page and can be removed
Run commands, hand tasks to agents, read and write files on your machines	Each time needs your approval in the Authorization Center, unless you set up an allow rule beforehand
Ask for your review and wait for your answer	Reply from the console or from Zhaoye

You can also limit it to specific machines.

Create a credential by hand

Some clients cannot request access on their own. Open Settings → Agent access in the console, choose Create manually, and pick permissions, machines and expiry. The credential is shown once, together with:

Claude Code and Codex: one command to copy and run;
Other clients: setup instructions you can hand straight to the client.

Treat the credential like a key. Put it in the client configuration, never in a chat, transcript or screenshot.

Revoke

Settings → Agent access lists every credential: which assistant it belongs to, what it may do and when it was last used. Choose Revoke and it stops working immediately.

Give each assistant on each machine its own credential so you can revoke one without affecting the rest.