Set up AI

Saiki talks to Claude and ChatGPT through the claude and codex command-line tools on your Mac, signed in with the plan you already pay for.

How it works

Saiki doesn't have its own AI and doesn't need an account or an API key. When you ask something, it runs the same command-line tool you'd run in Terminal, and that tool answers with your own subscription:

In Saiki	Runs	Signed in with
Claude	Claude Code (`claude`)	Your Claude plan
ChatGPT	Codex CLI (`codex`)	Your ChatGPT plan

It doesn't replace the Claude or ChatGPT apps. It's a quick way to reach the same AI from the bar, and any chat can move on to those apps or to Terminal when you want them (see Continue somewhere else). You need only one of the two tools; with both, you can switch between them in the middle of a chat.

Install and sign in

Install the tool in Terminal.
- Claude Code: curl -fsSL https://claude.ai/install.sh | bash
- Codex CLI: npm install -g @openai/codex
Sign in once, in Terminal: run claude for Claude Code, or codex login for Codex, and follow its steps.
Open the bar, type a question and press ⇥. See Asking and chatting.

Saiki finds the tools in ~/.local/bin, /opt/homebrew/bin, /usr/local/bin or anywhere on your login shell's PATH. There's nothing to point it at.

Setup help in the bar

If you open a chat with an AI whose tool isn't installed, the chat shows a card instead, such as Claude needs Claude Code. It has the install command, Copy Install Command and Open Setup Guide, which opens the tool's own setup page.

If the tool is installed but not signed in, the answer says so in plain words, for example: "Claude Code isn't signed in. Run claude once in Terminal, then ask again."

Only one installed? If your default AI's tool is missing and the other one is installed, ⇥ asks the one you have.

Staying signed in

Claude Code's sign-in lasts several hours. While a Claude chat is open, Saiki runs claude auth status every 3 hours, so Claude Code can refresh its own sign-in and a long-running chat doesn't suddenly say it isn't logged in. Saiki never reads your sign-in itself.

Light when idle

Nothing of the AI runs until you use it. A tool starts when you open a chat and stops after 5 minutes with nothing to do; your next message starts it again, and the chat carries on where it was. Turning AI off stops them straight away.

Update reminders

When a newer Claude Code or Codex is out, a quiet line appears under the chat's header while the AI isn't answering, such as "Claude Code 2.1.290 is available". It never pops up and never shows in the middle of an answer.

Update asks once more, showing your version and the new one, then Update Now runs the tool's own claude update or codex update. A tool that's idle restarts on the new version with your next message.
Saiki updates only tools installed with their own installer. If you installed one with Homebrew or npm, the line says "Update it with Homebrew" (or npm) instead.
The ✕ button hides the reminder until the next version comes out.
If your tool is too old for Saiki, or for the model the chat uses, the line becomes a warning saying which version it needs.

Claude Code's newest version is checked once a day; Codex's comes from the file Codex keeps itself. The installed versions are also listed on AI Usage.

Settings

Open Settings and choose AI & Agents. The switch at the top turns AI on or off.

Chats

Row	What it does
Default AI	Who answers when you press `⇥` to ask: Claude or ChatGPT. Type `claude` or `gpt` and `⇥` to pick one each time.
Claude mode	What new Claude chats may do without asking. See Approvals and modes.
ChatGPT (Codex) mode	The same for new ChatGPT chats.
Claude model	The model new Claude chats use. "Same as Claude Code" uses whatever the tool itself uses.
ChatGPT (Codex) model	The model new ChatGPT chats use, from the list your account offers.
⇥ asks AI	When `⇥` has nothing else to do, it asks your default AI about what you typed. Off: no hint, and Ask stays a row you pick.

The Agents card on the same page has the needs-you card, Codex terminal titles and the Claude Code hook. They're explained in Agents.

Permissions

AI asks for Notifications (so an agent that needs you can tell you when the bar is closed), and Automation for Terminal or iTerm the first time you jump to an agent's tab. Each is asked for only when it's needed. See Install and permissions.

Turning AI off

Turn off the switch at the top of AI & Agents, or AI on the Features page. Ask AI, the Agents view, the needs-you card and AI Usage leave the bar, the tools stop, and nothing of the AI loads until you turn it back on. A reply that's still being written finishes first.

The Claude Code hook stays in place. Turning AI off stops Saiki answering Claude Code, but the hook it added (if you turned it on) stays in Claude Code's settings and steps aside while Saiki isn't answering. To remove it, turn off Answer Claude Code from Saiki first. See The Claude Code hook.

What Saiki never does

It never reads your Claude or ChatGPT sign-in or tokens. Usage comes from the tools' own reports.
It never runs a tool in a mode that skips permissions, such as Claude Code's bypass mode or Codex's full access.
It never edits Claude Code's or Codex's settings files, except two things you confirm on a card: saving your default mode to the tool, and the Claude Code hook. Each changes only its own entry.
It keeps its own copy of your chats, in its own data folder.