Agents
The Agents view lists every AI agent session on your Mac, in Saiki, your terminals, the Claude app or the Codex app, with the ones that need you first.
Open the Agents view
- Type
agents(oragent,runs) and press ⇥, or run the Agents command. - Choose Show Agents in Saiki's menu bar menu.
- In a chat, ⌘D opens that chat in the Detailed level.
Type to filter by title, folder, agent or where it runs. While you type, past sessions that match show too, under Earlier.
What it shows
Agents are grouped by whose turn it is, with a summary on top ("1 needs you · 2 working", or "All quiet"):
| Group | Means |
|---|---|
| Needs You | Waiting on you: "Waiting for your OK", "Has a question", "Plan ready" or "Stopped on an error". |
| Working | Busy right now. |
| Done | Finished recently. |
| Earlier | Older sessions your filter found. |
Rows don't jump around while agents work: they're ordered by when you last wrote to them.
Every agent on your Mac
- Saiki's own chats, with Claude or ChatGPT.
- Claude Code sessions in any terminal, and the Claude app's sessions.
- Codex sessions, including the ChatGPT app's, read from Codex's background server when it's running (Saiki never starts it) and from Codex's own session files.
- Other agent tools running in a terminal, found by their process with no permission needed: Gemini CLI, Cursor Agent, OpenCode, Copilot CLI, Grok, Amp, Kimi, Qwen Code, Pi, Droid, Goose, Aider, Antigravity, Kiro, Cline and Hermes. In progress
Simple, Detailed and Board
| Level | What you see | Keys |
|---|---|---|
| Simple | A list at the bar's usual size. | ⌘D to switch to Detailed |
| Detailed | A bigger bar: the list on the left, the selected agent's conversation and changes on the right. | ⌘D back to Simple |
| Board | Detailed as columns, one per group. | ⌘B switches List and Board |
↑ ↓ move; on the Board, ← → move between columns. In Simple and on the Board, Space or ⌘I opens the selected agent's conversation in place, and esc goes back. Saiki remembers the level and layout you used last.
Jump, Open or Continue Here
↵ takes you to the agent where it lives, and the footer says where. ⌘K lists every way that applies.
| Action | What happens |
|---|---|
| Open Chat | For Saiki's own chats: opens the chat in the bar. |
| Jump to Terminal | For an agent running in a terminal (the action names the app, such as iTerm or Ghostty): brings its tab to the front. That's the exact tab in Terminal and iTerm (the first time, macOS asks to let Saiki control them), the pane in tmux, and the app in any other terminal. |
| Open in Claude | Opens the session in the Claude app. |
| Open in Codex | Opens the session in the Codex app. |
| Continue Here | For an ended Claude Code or Codex session: turns it into a Saiki chat that resumes the same session in its folder, with the conversation so far, ready to type. |
| Continue in Terminal | For an ended session: resumes it with the full tool in a new terminal window. |
| Show Details | Shows its card (⌘I). |
Saiki never writes to a session something else is running. A running agent is answered where it runs: the view says "Running in Terminal · reply there" and jumps you to it. A session open in another app can continue here about a minute after you close it there. Continuing in a folder Saiki hasn't run an agent in before asks you to trust it first (see Folder trust).
Find past sessions from search
You don't need to open the Agents view to find an old session: type part of its title in the bar and it shows among the results. ↵ opens it where it lives, or in the Agents view; ⌘K also has Show in Agents, Copy Session ID and Show Folder in Finder.
Talk to a chat in Detailed
In Detailed, Saiki's own chats and ended sessions get a message box under the conversation. Press ⇥ to go into it ("Message Claude…"), ↵ to send, ⇧↵ or ⌥↵ for a new line, and esc to go back to the list. An ended session is continued here first. The / menu and attaching files work as in the bar chat.
Under the box, chips change the chat's settings:
| Chip | What it changes |
|---|---|
| Model ⌘⇧M | The model; picking the other AI's model carries the chat on with it. |
| Traits ⌘⇧E | Effort, speed (Fast uses your limits faster) and, for Claude models that offer it, a 200K or 1M context. |
| Access ⌘⇧A | The permission mode. See Approvals and modes. |
| Plan | Plan first: the agent changes nothing until you implement its plan. |
Your effort and speed picks are remembered for the next chats with that model. The same choices are in ⌘K as submenus.
Changes
⌘⇧D, or Show Changes in ⌘K, shows what a turn changed, in place of the conversation: the files (↑ ↓) and each file's diff with line numbers (⌥↑ ⌥↓ between changes). esc goes back.
- For Saiki's chats, Revert Changes and Undo Revert are here too.
- For Claude Code sessions started elsewhere, you see the edits Claude Code recorded. Edits it made through shell commands aren't in those records, and these can't be reverted from Saiki.
Answering from the Agents view
When the selected agent is waiting on your OK, its ask is shown in full: ⌘↵ allows it and ⌘⌫ denies it. For a question, ↑ ↓ move through its options, ↵ picks, ← → change question and ⌘↵ sends. This works for Saiki's own chats, and for Claude Code anywhere once the Claude Code hook is on.
An answer only ever applies to the ask you were shown. If the agent has moved on to another, nothing is answered and you see the new one. An ask that just appeared can't be allowed in its first moment ("Take a look first"), so a key meant for something else never approves it. An ask too long to show whole can only be denied here, or allowed in the terminal.
The needs-you card
When an agent asks for your OK or asks you a question and you aren't looking at it, a small glass card slides in at the top centre of the screen, over whatever app you're in. It shows the agent, its folder and the exact ask, such as the command it wants to run or the file it wants to edit.
| Keys | Button | What it does |
|---|---|---|
| ⌃⌥↵ | Allow | Allows it, where Saiki can answer. |
| ⌃⌥⌫ | Deny | Denies it (Skip, for a question). |
| ⌃⌥O | Open | Goes where the agent lives: its terminal tab ("Open · Terminal"), the Claude or Codex app, or the bar for Saiki's chats. |
| ⌃⌥↑ ⌃⌥↓ | For a question: move through its options; ⌃⌥↵ picks. |
The keys work from any app, and the card never takes the keyboard from the app you're typing in. Only real key presses and clicks count, so an agent driving the keyboard or mouse can't approve its own ask, and Allow waits until the card has settled.
- After about 3 seconds the card tucks into a small pill ("Claude needs you", or "2 waiting"). Point at it to open it again; its ✕ puts it away until something new needs you.
- It goes by itself once the ask is answered anywhere, in the terminal included.
- Allow shows for Saiki's own chats and for Claude Code through the hook. Other agents get Open, so you answer them where they run.
- A question with one answer is answered on the card; questions with several open the Agents view.
- A plan to review or an agent stopped on an error stays in Needs You without a card.
When you're away, the screen is locked, an app is full screen, or the card is turned off, you get a macOS notification instead. It offers Open (Saiki's chats also get Deny, and Reply for a single question), never Allow.
The bar never jumps to an agent by itself. When you next open it, agents waiting on you are its first rows, with ↵ Review opening them in the Agents view. In progress
Turn the card off with Show a card when an agent needs you in Settings → AI & Agents.
Codex in a terminal
Codex in a terminal shows that it needs you only in its window title ("Action Required"). Turn on Know when Codex in a terminal needs you in Settings → AI & Agents, and Saiki reads the titles of the terminal windows where Codex runs (never what's on screen), so the card comes up with Open to jump to it. It's off by default. Saiki can't answer these from the card; you answer in the terminal.
Agents in the menu bar
Saiki's menu bar icon gets an orange dot while an agent needs you; point at it to see how many need you and how many are working. Its menu lists up to eight agents (those that need you first) with what each is doing; choose one to open the bar at it, or Show Agents for the whole view. The menu also has Keep Awake. See Menu bar.
The Claude Code hook
Claude Code asks for permission in its own terminal. To answer those asks from Saiki too (in the Agents view and on the needs-you card), Saiki can add one hook to Claude Code's settings. It's off until you turn it on:
- The Agents view offers it once, as Answer Claude from here: Turn On or Not Now.
- Or turn on Answer Claude Code from Saiki in Settings → AI & Agents. A card says exactly what changes before anything does: Add the Hook, or Remove the Hook when you turn it off.
What it changes
- It adds one
PermissionRequesthook to~/.claude/settings.json, pointing at a small helper inside Saiki. Every other line of the file stays exactly as it was, and Saiki keeps a backup of the file first. - The terminal still asks as usual. Whichever you answer first, in the terminal or in Saiki, wins, and the other goes away.
- If Saiki isn't running or doesn't answer, the helper steps aside and Claude Code carries on as if it weren't there.
- Sessions in the Claude app are answered in the app; Open on the card takes you there.
- Turning it off removes exactly that entry and nothing else. If you remove it yourself (for example with Claude Code's
/hooks), Saiki never adds it back on its own.
This hook is the only change Saiki makes to Claude Code's settings for sessions you start elsewhere, and only with your OK.