Skip to content

Tools in the app

Fewer tools make a better agent — every tool is one more thing a small model can pick wrongly — so the model sees a small everyday set, and the rest wait in Settings, off until you want them. Every tool's schema was measured on real 0.8B and 2B models before it shipped; a tool this phone can't run (no Image Playground, no translation pack) is hidden from the model rather than offered and failing.

Every call is a row: arguments, result, time.
Writes ask first.
Record a take; the model calls transcribe.
current_time · calculator any time zone; exact arithmetic
remember · recall notes the agent keeps between chats, on the device
http any HTTPS request — writes ask first; refused in offline mode
device_info battery, storage, model, OS
manage_messages the model tidies its own context in a long chat — compact drops old tool calls and results but keeps every message's text, drop/clear remove old turns, stats/list show the size. The chat on screen never changes; the transcript gets one quiet line, "2 turns compacted by the agent · your history is unchanged". No approval card — it edits only what the model sees
system_prompt the model reads or changes its own system prompt when you ask for a standing rule — "from now on always answer in Turkish" → append; "what are your instructions?" → get; set replaces, reset restores the default. Every change stops at an approval card that shows old → new; after you approve, it applies from the model's very next step and to every new conversation. Settings › System prompt then reads changed by the agent · 8:25 PM with one-tap Revert (the last five prompts are kept)
render_ui a native card instead of prose — bar/line/pie charts (Swift Charts), a map with pins and a route (tap opens Maps), a flow diagram, tables, key/value grids, lists, progress bars, markdown. The model sends a small typed spec; the app draws it — nothing runs, no web view. An unknown component type is named on the card, never dropped. Cards stay in History and re-render from the saved call
ask_user the model asks you mid-turn with a native card — Yes/No, a radio list, checkboxes, a text field, a row of chips, a small form (text · password · number · toggle), or a note to acknowledge. The agent waits (180 s by default) with a visible waiting for you state; Cancel is always there and the model is told honestly. Passwords are typed into a secure field and never appear in the chat, History or an export — only the model gets them, once, to use them. How it works
generate_image a picture from a prompt (Image Playground)
generate_voice speaks with on-device voices, Personal Voice included
transcribe speech to text, offline — tap the mic, record, send; the model sees [audio attached: …] and decides to call it
translate Apple's offline translation packs
describe_image scene labels, faces, barcodes on an attached photo (Vision)
flashlight · vibrate torch on/off/dimmed; haptic patterns
open_app another app via deep link, with context in the query string
music play, pause, next, previous, volume and what's playing for the system music player (Apple Music or whatever holds it) — one tool, no approval card; the volume goes through iOS's own slider, so the system volume HUD appears
clipboard_read · clipboard_write what you copied — text, a link, or a picture the model can describe_image — and, after asking, text copied back. Two tools, not one with an action: measured on a 0.8B, "what did I copy?" turned into a write 4 times in 6 with the single tool

Off by default

Tools you need now and then stay off until you switch them on in Settings › Tools — every extra tool is one more thing a small model can pick wrongly, and the measured difference is real. Each one asks for its iOS permission the first time it runs, not at install.

calendar_events · calendar_create what is on your calendar for the coming days; adding an event asks first (EventKit)
reminders_due · reminders_create open reminders due today or soon; adding one asks first
location your city and coordinates rounded to about 1 km, plus the time zone — Exact position is a switch in Settings, never something the model can ask for (measured: a 0.8B asked for it every single time); purpose: weather hands the model the forecast URL for http
contacts_search a name → phone numbers and emails, read-only; the model chains it into open_app with tel: / sms: / mailto:
photos_recent your newest photos, or one day's, handed to the model as pictures — Selected Photos access is enough, so you decide which pictures loopl may ever see

Reads and creates are separate tools on purpose: with one tool carrying both verbs a 0.8B model turned "what did I copy?" into a write 4 times in 6. The create tools' descriptions carry today's date, so "tomorrow at 12:30" becomes a real timestamp without a second tool call.

The model can mount a tool itself

manage_tools is always on. When the model needs a tool that is switched off — say you asked it to fetch a page while http is off — it can call manage_tools list, see every tool the app has (and why one cannot run on this device), and mount the one it needs. The mount counts for this conversation and works on the model's very next step, inside the same answer; a new conversation starts from your switches again. A small model often skips the ceremony and just calls the tool it wants — that works too: calling a known, switched-off tool mounts it for the conversation and runs it, and the result says so. Settings › Tools shows the difference as mounted for this conversation / unmounted for this conversation and one line under the list says what the agent changed. Your switch always wins: flipping a tool in Settings applies to the current conversation too.

Measured, not assumed: the ManageToolsMeasurement test runs the real loopl 0.8B and 2B models with the needed tool unmounted and records whether the model mounted it and used it in the same turn (numbers in the changelog).

Asking first

A call that writes — http with POST/PUT/DELETE, open_app, clipboard_write — stops at an approval card: the exact request, Approve, Deny, or Edit the arguments. Deny, and the model is told it was denied and carries on without it. Settings › Network › Ask before chooses: every request, writes only (the default), or never. An unanswered card is denied after two minutes.

open_app asks every time unless you put the app's scheme on its allowlist (Settings › Tools › open_app › Open without asking). shortcuts:// is the useful one: the agent can run a Shortcut with text as input.

Controls on a result

Some results come with a live control right under the tool row — no need to ask the model again:

Result Control
flashlight on Turn off (and Turn on once it's off)
vibrate Vibrate again
music a mini player — previous · pause/play · next, and the volume slider
location a pin on a small map; tap to open Maps
calendar_create · reminders_create Open in Calendar / Open in Reminders

Tapping one runs the same tool with those arguments through the normal path — if the tool asks first, you get the usual approval card — and the row updates with the fresh result ("flashlight: off", the switch now says Turn on) plus a quiet updated by your tap. It is not a new message: the row keeps its original arguments and no new row appears; History saves the latest state; the model is told the user tapped this in the app so its next answer is right. If the tap is denied or fails, the card keeps its last true state and the error line under the chat says why. Controls are drawn disabled while an answer is being generated or when no model is loaded — the state they show is still true. VoiceOver reads the state ("Turn off, currently on").

When a call fails

The model sees the error, word for word, and tries again with better arguments or answers without the tool. loopl never silently retries or blocks a repeat — what you see in the row is what the model read.

Building your own? The same tools are in the SDK: Tools.