Asking and chatting

Type a question in the bar and press ⇥: the answer streams in right there, and the field becomes the box for your follow-up.

Type, press ⇥, and Claude answers in the bar.

Chats run on your own Claude or ChatGPT plan, through the claude and codex tools on your Mac. If you haven't set those up yet, start with Set up AI.

Ask from the bar

Type	What happens
A question, then `⇥`	A new chat with your default AI about what you typed. The bar shows "Ask AI ⇥" when `⇥` has nothing else to do.
`ai` `⇥`	Opens your current chat with the default AI, or an empty one. `ai` is the Ask AI command's alias; `ask` works too.
`claude` `⇥`	The same, with Claude.
`gpt`, `chatgpt` or `codex` `⇥`	The same, with ChatGPT.
Anything, then the last row	Every search ends with "Ask Claude about “…”" (or ChatGPT, whichever is your default), next to the web search row.

⇥ does other things first when they apply, like completing a suggestion, a path or a site's search shortcut; it asks the AI only when there's nothing else to do. You can turn that off with ⇥ asks AI in Settings → AI & Agents, and pick who answers with Default AI.

The chat

Your messages sit on the right, answers on the left. Answers are formatted (headings, lists, code blocks), and while the AI works you see what it's doing ("Thinking…" or the step it's on). Its thinking folds into a "Thought for 6s" line you can open.

The header shows the AI and the model it's running, the chat's permission mode, and a small meter for context and usage (see Usage and context). Point at an answer to see Copy, Ask Again on the newest one, and the time.

Keys in a chat

Keys	What they do
`↵`	Send. With the field empty: copy the last answer.
`⌘↵`	Paste the last answer into the app you were in (the bar closes).
`⌘R`	Ask Again: the last question once more.
`⌘N`	New chat.
`⌘[` `⌘]`	Previous and next chat.
`⌘1` `⌘2`	Carry this chat on with Claude or ChatGPT.
`⌘P`	The next model.
`⇧⇥`	The next permission mode. See Approvals and modes.
`⌘U`	Attach files.
`→`	With the field empty: open or close the newest answer's steps.
`⌘D`	Open this chat in the Agents view's Detailed level.
`⌘K`	Everything else you can do with the chat.
`esc`	Leave the chat. It's kept: `ai` `⇥` brings it back, and `⌘[` `⌘]` step through older chats from inside one.

⌫ on an empty field never closes a chat, so a stray key can't throw it away. Chats are saved as you go and are still there after Saiki restarts.

While it answers

Stop

Press ⌃C (or ⌘.), or click Stop in the footer. The answer keeps what it had written, marked Stopped, and the field asks "Tell Claude what to do instead…", so you can say what you want next.

Queue a message

You can keep typing while it works. ↵ queues your message: it waits above the field, marked Queued, and goes by itself when the answer ends. Several queued messages go together as one.

↑ on an empty field takes the newest queued message back to edit it.
⌘⌫ cancels it.
⌃C with messages queued stops the answer and sends them straight away.

Send Now

⌘↵ while it works hands your message to the AI without stopping it ("Sent, waiting for Claude"). It takes it at its next step, and the answer carries on from there with your words in mind. With the field empty, ⌘↵ sends what's queued. If the answer has just ended, the message goes back in the queue and is sent next.

Switch between Claude and ChatGPT

In a chat, ⌘1 carries it on with Claude and ⌘2 with ChatGPT (also Continue with… in ⌘K). It's the same conversation: with your next message, the other AI is told what was said while it wasn't there. The new AI starts in your usual mode and default model for it. You need both tools installed.

Choose a model

⌘P steps through the models your AI offers.
Type /model and a space for the list; ↑ ↓ and ↵ or ⇥ pick one. /model followed by a name picks it directly.
The model for new chats is Claude model and ChatGPT (Codex) model in Settings → AI & Agents. The lists are what each tool reports, so they're the models your account can use.

Effort, speed and context size are on the Traits chip in the Agents view's Detailed level. See Talk to a chat in Detailed.

Images and files

Files you add show as chips in the field and go with your next message. If you send only files, the question is "What's in this file?".

Paste: ⌘V with files copied in Finder, or an image on the clipboard (a screenshot, an image copied from a browser). A pasted image is saved as a PNG in Saiki's own data folder.
Drop: drag files onto the bar, in a chat or on the empty bar. A card shows what will happen; when File Converter is on, it has two halves, and dropping on Ask AI attaches the files.
Pick: ⌘U or Attach Files… in ⌘K opens a file picker; you can pick folders too.
From search: on a file or folder in the results, ⌘U is Ask AI About This File. ⌘-click several results, then ⌘K → Ask AI About These Files.

Images go to the AI as images; other files are passed by their path, for the AI to read. ⌫ on an empty field takes the last chip back out. Apps can't be attached.

Slash commands

Type / at the start of the field for a menu of commands; keep typing to filter, ↑ ↓ to move, ↵ or ⇥ to pick, esc to hide it. Saiki's own commands come first:

Command	What it does
`/clear`	Start a fresh chat.
`/compact`	Summarize the chat so far to free context (uses your limits).
`/model`	Switch this chat's model.
`/status`	Context and usage for this chat.
`/rename`	Give this chat a title.
`/export`	Save the chat as a ZIP in Downloads, with the chat and a readable transcript.
`/review`	Review the uncommitted changes in the chat's folder, or the changes against a branch you name. It doesn't change anything.
`/plan`	Plan before changing anything (Claude).
`/default`	Leave Plan mode, back to your usual mode (Claude).
`/debug`	Debug mode on or off: reproduce, test one idea at a time, verify.
`/subagents`	Ask it to split the task across subagents.

After those come the AI's own: Claude Code's commands and skills (including your custom ones), and ChatGPT's skills. When a name is taken by both, Saiki's command wins, since it acts on Saiki's copy of the chat.

Continue somewhere else

Any chat can carry on outside the bar, as the same conversation. Saiki lets go of it first, so only one place writes to it at a time.

Keys	In ⌘K	What happens
`⇧⌘O`	Continue in Claude App / ChatGPT App	Opens the chat in the Claude or ChatGPT app. Shown only when that app is installed.
`⇧⌘T`	Continue in Terminal	Opens your terminal app with the full `claude` or `codex` tool, resuming this chat in its folder.

A Claude chat can continue elsewhere once you've sent a first message; a ChatGPT chat once ChatGPT has answered in it.

More in ⌘K

⌘K in a chat lists everything above with its keys: Copy Answer, Paste Answer, Ask Again, Stop, Send Queued Now, Cancel Queued Message, Revert Changes, Continue with…, Attach Files…, Change Model, Change Mode, Open in Detailed View, New Chat, Previous Chat and Next Chat. It also has Restart Agent, so the chat picks up skills, plugins and MCP servers you added since it started: a Claude chat's tool stops and resumes the same chat with your next message, and ChatGPT reloads its commands.

Where chats run

A new chat works in Saiki's own Workspace folder, so asking an AI to "create a file" doesn't scatter files around your home folder. The AI can still read and work elsewhere when you ask, under the chat's permission mode. A chat continued from another app keeps its own folder, and the header names it.

Saiki keeps its own copy of every chat, so a chat survives switching AI, and you can find it again from search or the Agents view.