All pages

Docs · More

UI reference — the Agent tab

Availability: macOS.

The AI chat lives in the right sidebar's Agent tab — the first segment of the strip at the top of the right sidebar (show the sidebar with ⇧⌘\). Labels below are copied from the source; quote them as written. Behaviour in depth: Agent, Agent safety, Scheduled runs.

Top to bottom the tab is: header row → chat tabs → meters (Context / cached / posture) → banners → transcript → spend footer → Snapshots row → chips → composer.

Header row

  • Model badge — short tag coloured by provider ("Active model: …").
  • Folder chip — the working directory the agent runs tools in; click to pick another folder (it cds the focused pane, or sets the CLI harness folder).
  • Stop button (red while running) — "Stop the running agent turn", ⌘..
  • Model picker (shows the model name). Sections:
    • "Your CLI subscription (no API credits)": Claude Code, Codex, Gemini CLI ("— not installed" when missing). Picking one runs that CLI as the agent instead of an API model.
    • One section per provider: Anthropic, OpenAI, Google Gemini, Ollama / Local, OpenRouter.
    • "Pay with": Your API key / TRMNLY credits (only when the model is offered on credits — see Cred wallet).
    • "New chats": Use as default for new chats (or "…in this project").
  • Gear — "Agent settings" menu (below).
  • Shield (top-right) — opens the Agent safety sheet. An orange dot means the sandbox blocked something.

Chat tabs

  • One tab per chat (default title "Chat"); click to switch, double-click to rename. + = "New chat". x = "Close tab" (when 2+ are open).
  • Right-click a tab: Rename, Stick to top / Unstick from top, Archive, Delete permanently….

Gear menu ("Agent settings"), in order

  • Auto-approve all tools (same flag as the Auto-accept posture).
  • Text size: Decrease text size, Increase text size, Reset text size; High-contrast colors.
  • Session: New chat ⇧⌘N, Next chat ⇧⌘], Previous chat ⇧⌘[, Search chats… ⇧⌘F, Archived chats…, Project agent settings…, File-grouped diffs….
  • Persona: Default or a named persona.
  • Pairing: Pairing mode, Add person….
  • Automation: Scheduled jobs… (opens the sheet titled "Scheduled runs"), Watch tasks…, Auto-ask agent on build failures.
  • Context: Scope (projects & tabs)…, Pinned files…, Memory…, Scratchpad…, What the model sees… ⇧⌘I, Fold tool calls in transcript, Project recall index (semantic_recall).
  • Providers & routing: Providers, routing & spend…, Auto-route small/large model, Escalate model on trouble, Self-critique multi-file edits, Anytime answers (fast provisional reply).
  • Mode: Full access, Dry run (echo commands), Observe (read-only).
  • CLI agent (only when the model is a CLI harness): Read-only, Accept edits (default), Full access (no sandbox); Bill to my CLI login, not API keys.
  • Observability: Turn traces…, Eval harness…, Tool telemetry….
  • Transcript: Copy as Markdown, Undo last write: file ("Undo last agent write"), Browse file snapshots…, Export as plain text…, Export as Markdown…, Export as JSON…, Share session…, Replay this session…, Compare with another session…, Save as Runbook…, Report a problem with this conversation….
  • Voice / input: Speak assistant replies, Voice backend (System voice / ElevenLabs), Configure ElevenLabs…, Vim mode for input.
  • Set API key…, Clear conversation.

There is no separate "Plan" or "Ask" mode in the Mac Agent tab: the modes are the gear menu's Mode section and the posture chip below.

Meters and posture

  • Context meter — "NN% · used / window"; orange past 75 %, red past 90 %.
  • "n cached" chip — prompt-cache hits on the last request.
  • "chat only" chip — a local model without tool support.
  • Posture chip — Approve (confirmation card before each sensitive tool) / Auto-accept (everything runs) / Observe (read-only; writes and commands blocked), or Dry-run. Click it, or press Shift-Tab in the composer, to cycle.

Banners

  • Missing key: "Add your … API key to start." (click opens the key sheet), or the variant offering to pair this Mac for credits.
  • Error row: Retry, Add key, Use credits, Pair this Mac, Use my API key depending on the error; right-click Copy error.
  • Cost-cap rows: Raise to …, Disable cap, Reset today, Disable daily cap.
  • "Agent locked down" with Audit / Resume after a panic (⇧⌘.).

Transcript

  • Empty state "Ask the agent for help." with three suggestion cards.
  • Jump to latest / Latest button when scrolled up.
  • Your message — hover: More actions (…), Edit and resend, Branch from here; right-click also Copy text, Replay with model…, Pin message, Add tag…. Edit opens "Edit this turn" → Save & resend.
  • A reply — hover: Copy reply as Markdown, Retry from the prompt, in a new branch, Branch from here; a "~N tok · $X" pill.
  • Tool card — status chip (approval needed / running / done / denied / failed / canceled); Cancel while running; right-click Why did you do that?, Copy arguments as JSON, Copy result, Copy tool name. Write/edit cards show a diff.
  • Pending approval — Approve (reads Approve edits after editing), Deny, Edit args, and Always… (Allow this tool, Allow this path, Allow this directory, Allow on host, Deny this tool, Deny on host). Two or more pending file changes add a batch bar with Approve all / Deny all. Irreversible git actions on a guarded branch need Confirm — step 1 and Confirm — step 2.
  • Keyboard (focus in the sidebar): ⌃J / ⌃K next / previous card, ⌥Space expand, ⇧⌘C copy card, ⌘↩ approve the focused card (otherwise send), ⌘⌫ deny.
  • "# N tok" (session tokens), session "$X.XX", "today $X" (daily spend, bar when a daily cap is set), provider badge and usage pill. For a CLI harness the cost is what the CLI reported, plus "plan" when on a subscription.
  • Snapshots row with Browse…; per file a revert button ("Restore … as it was before this write" or "Delete …") and a compare button. This is how a user undoes a file change the agent made.

Composer

  • Placeholder "Message the agent…". ↩ sends, ⇧↩ newline, ↑/↓ in an empty field recalls earlier prompts, / opens the slash palette, ⇧⌘V pastes as plain text. ⌘↩ always sends (queues if a turn is running).
  • Scope chip ("Project", "Current", "All projects", "+N", "· N tabs") — "Choose which projects and tabs the agent uses as context". Popover: "Projects in context" (Current / All / Pick…), "Extra projects", "Include" (Pinned files, Project README / docs, Memory index, System-prompt addendum, Project notes), "Attach tabs".
  • Queued prompts strip ("N queued prompts") with Resume, Clear, scope Tab / Project.
  • Button row, left to right: vim pill; mic ("Dictate prompt"); hands-free (⌃⌘I); external editor ("Compose in external editor"); paperclip ("Attach a file to the message"); globe (web search on/off for this chat); camera ("Capture a screenshot into the chat"); bookmark ("Save the current draft as a reusable prompt"); Send ("Send (⌘⏎)") / Stop ("Stop generating (⌘.)").
  • Smart chips once the chat has messages: "fix last error", "explain last command", "review diff".
  • Slash palette: built-ins /test, /review, /diff, /explain, /fix, /summarize, /search, then your saved prompts; footer "Save current draft as prompt…".

Sheets opened from this tab

  • Search conversations — "Limit to current project", "Include archived".
  • Archived chats — Resume per chat.
  • Project agent settings — Mode, "System prompt addendum", "Include project context (CLAUDE.md, README.md, …)", "Cost cap (USD)", "Summarize at", "Tool allowlist", "Saved approval rules".
  • What the model sees (⇧⌘I) — tool definitions and the transcript as sent.
  • Pinned files, Agent memory (New entry, Save ⌘S, Delete), Conversation scratchpad, Agent context (scope, with Preview).
  • Providers, routing & spend — API keys (Anthropic / OpenAI / Gemini / OpenRouter), "Per-session model routing" (Default for new chats, Tool model, Auto-route small model, Fallback provider), "Extended thinking (Anthropic)", "Daily spend cap", "Context window", "Local (Ollama / LM Studio)" endpoint / model, "OpenRouter" model id, "Smart-action palette" global shortcut.
  • Scheduled runs (gear → Scheduled jobs…) — Pause all, Add; per row Run, Edit, delete; editor: Name, target "Agent prompt" / "Terminal command", "Schedule" (e.g. @daily 09:00), timezone, quiet hours, days, catch-up and idle options.
  • Watch-mode tasks (gear → Watch tasks…) — path to watch, prompt, debounce.
  • API key sheet — "Provider API key", Reset Keychain entry, Save.
  • Agent safety (shield) — tabs Audit Log, Allowed Folders, Command Rules, Tools, Network, Detectors, Session; button Panic — lock down / Resume agent.