Capabilities

Ask a target what works there before your agent relies on it. One command answers with a snapshot: each capability is available, degraded or unavailable, and anything short of available says why.

Try it autonom capabilities

Needs Autonom installed · a ready emulator, Android phone or Simulator

How it works

autonom capabilities describes a target, not a session. With a session it describes the session's target and fills in session_id. Without one it takes the target your flags name, or the only ready target, and session_id is null. That snapshot assumes nothing is attached yet, so network.capture is unavailable until a session attaches the capture proxy.

The answer is an autonom.provider/v1 snapshot inside the usual ok envelope. provider is local.android or local.ios. device_class is simulator for an Android emulator or an iOS Simulator, and physical for an Android phone. captured_at says when it was taken. capabilities lists every capability by name, in alphabetical order.

StateWhat it means
availableWorks on this target now.
degradedWorks in a weaker form. reason says how, for example a checkpoint that restores by replaying the flow from its start.
unavailableDoes not work here. reason says why, for example a control that needs an emulator or Simulator.
unknownAllowed by the schema for a provider that cannot tell. The local Android and iOS providers in v0.36.1 always answer with one of the other three.

A snapshot is a record, not a promise: it never changes after captured_at. A flow can name the capabilities it needs under requires, and flow run takes a fresh snapshot and checks them before the first step. A missing one stops the run with flow_requirements_unmet and nothing touches the device.

--probe adds a measurement. It reads the current screen once, so on iOS ui.accessibility and ui.input describe what was measured, not only which tools are installed. On Android the tree and input come from adb, so the answer is the same with or without it.

Examples

Every output here comes from Autonom's test emulator, a test Android phone and a test Simulator, so you can rerun it without a device. Reproduce these outputs, at the end of this section, has the exact commands.

See what an emulator can do

Goal: learn what your agent can rely on before it starts a long check.

  1. Ask the emulator.

    No session is needed; --serial names the target.

    autonom capabilities --serial emulator-5554

You should see the snapshot, with each capability and its state.

{ "ok": true, "session_id": null, "schema": "autonom.provider/v1", "provider": "local.android", "target_id": "emulator-5554", "device_class": "simulator", "captured_at": "2026-10-19T06:45:26.128Z", "capabilities": [ { "name": "checkpoint.create", "state": "degraded", "reason": "portable flow-prefix checkpoint" }, … { "name": "network.capture", "state": "unavailable", "reason": "capture proxy is not attached" }, … ]}

What to look at. device_class is simulator, because an emulator can be reset and controlled from the host. checkpoint.create is degraded: a checkpoint restores by replaying the flow from its start. network.capture waits for a session to attach the capture proxy.

Show full output
{ "ok": true, "session_id": null, "schema": "autonom.provider/v1", "provider": "local.android", "target_id": "emulator-5554", "device_class": "simulator", "captured_at": "2026-10-19T06:45:26.128Z", "capabilities": [ { "name": "checkpoint.create", "state": "degraded", "reason": "portable flow-prefix checkpoint" }, { "name": "checkpoint.restore", "state": "degraded", "reason": "replay from flow start" }, { "name": "logs", "state": "available" }, { "name": "network.capture", "state": "unavailable", "reason": "capture proxy is not attached" }, { "name": "screen.stream", "state": "available" }, { "name": "screenshots", "state": "available" }, { "name": "simulator.appearance", "state": "available" }, { "name": "simulator.battery", "state": "available" }, { "name": "simulator.biometric", "state": "available" }, { "name": "simulator.call", "state": "available" }, { "name": "simulator.clipboard", "state": "available" }, { "name": "simulator.keyboard", "state": "unavailable", "reason": "Android keeps keyboard preferences inside the keyboard app; no host-level store reaches them" }, { "name": "simulator.location", "state": "available" }, { "name": "simulator.network", "state": "available" }, { "name": "simulator.permissions", "state": "available" }, { "name": "simulator.push", "state": "unavailable", "reason": "the platform exposes no provider-neutral control" }, { "name": "simulator.sms", "state": "available" }, { "name": "simulator.status_bar", "state": "available" }, { "name": "simulator.text_size", "state": "available" }, { "name": "ui.accessibility", "state": "available" }, { "name": "ui.input", "state": "available" } ]}

Compare it with an Android phone

Goal: see what changes when the target is a phone instead of an emulator.

  1. Ask the phone.

    The serial is the one adb devices lists for it.

    autonom capabilities --serial R5CT20ABCDE

You should see device_class change, and the emulator-only controls become unavailable.

{ "ok": true, "session_id": null, "schema": "autonom.provider/v1", "provider": "local.android", "target_id": "R5CT20ABCDE", "device_class": "physical", … "capabilities": [ { "name": "checkpoint.create", "state": "unavailable", "reason": "physical devices cannot be snapshotted" }, { "name": "checkpoint.restore", "state": "unavailable", "reason": "physical devices require baseline replay" }, … { "name": "simulator.location", "state": "unavailable", "reason": "requires an emulator or simulator" }, … { "name": "ui.accessibility", "state": "available" }, { "name": "ui.input", "state": "available" } ]}

What to look at. The screen, input, screenshots and logs work the same. Every simulator.* control, such as a simulated location, needs an emulator. A phone cannot be snapshotted, so a replay always starts from the beginning of the flow.

Measure the screen with --probe

Goal: tell an installed UI backend from one that can read this screen.

  1. Ask the Simulator without measuring.

    The answer reflects the tools that are installed.

    autonom capabilities --udid AAAAAAAA-1111-2222-3333-BBBBBBBBBBBB
  2. Ask again with --probe.

    Autonom reads the current screen once before it answers.

    autonom capabilities --udid AAAAAAAA-1111-2222-3333-BBBBBBBBBBBB --probe

You should see ui.accessibility available first, then degraded once the screen was read.

{ "ok": true, "session_id": null, "schema": "autonom.provider/v1", "provider": "local.ios", "target_id": "AAAAAAAA-1111-2222-3333-BBBBBBBBBBBB", "device_class": "simulator", … { "name": "ui.accessibility", "state": "available" }, …}
{ "ok": true, … { "name": "ui.accessibility", "state": "degraded", "reason": "empty_accessibility_tree" }, { "name": "ui.input", "state": "available" } ]}

What to look at. The idb backend is installed, so the first answer says available. The measured tree was empty, so --probe says degraded with the reason empty_accessibility_tree. On a real app that usually means its views carry no accessibility labels; Read and drive the screen explains what to do next.

Name the target when several are ready

Goal: see what your agent gets when it asks without naming a target.

  1. Ask with no target flags.

    An emulator, a phone and a Simulator are ready, so Autonom will not pick one.

    autonom capabilities

You should see one error object on stderr, with exit code 2.

{"ok": false, "error_code": "ambiguous_target", "error": "3 ready targets; pass --target (or --platform)", "hint": "Candidates: android:emulator-5554, android:R5CT20ABCDE, ios:AAAAAAAA-1111-2222-3333-BBBBBBBBBBBB", "candidates": [{"platform": "android", "target_id": "emulator-5554"}, {"platform": "android", "target_id": "R5CT20ABCDE"}, {"platform": "ios", "target_id": "AAAAAAAA-1111-2222-3333-BBBBBBBBBBBB"}]}

What to look at. ambiguous_target lists the candidates. Pass one with --target, --serial or --udid, or narrow the list with --platform. Once a session owns a target, plain autonom capabilities describes that target.

Let a flow refuse a target that lacks a capability

Goal: stop a flow before its first step when the target cannot do what the flow needs.

  1. Name the capability in the flow.

    This flow checks requests, so it needs the capture proxy.

    schema: autonom.dev/flow/v1appId: com.example.appname: Checkout requestsrequires: capabilities: [network.capture]---- launchApp
  2. Run it in a session without the proxy.autonom flow run checkout.yaml

You should see the run refused with exit code 2, before the app launches.

{"ok": false, "error_code": "flow_requirements_unmet", "error": "active provider does not satisfy required capabilities: network.capture", "hint": "Inspect the immutable capability snapshot and select a compatible target.", "missing_capabilities": ["network.capture"], "capability_details": [{"name": "network.capture", "state": "unavailable", "reason": "capture proxy is not attached"}]}

What to look at. capability_details carries the snapshot entry for each missing capability, reason included. Attach the proxy with autonom network attach --i-understand-mitm, or run the flow on a target that has what it needs.

Reproduce these outputs

From the root of an Autonom checkout at tag v0.36.1 (commit ca2a280), in bash or zsh. The autonom function runs that checkout against its test doubles for adb, simctl and idb, so it needs no device, Xcode or idb. Session ids and times differ on every run.

# autonom below runs this checkout against Autonom's test emulator, an Android phone and a booted test Simulator.S=$PWD W=/tmp/autonom-capabilitiesrm -rf $W && mkdir -p $Wexport AUTONOM_HOME=$W/.autonom AUTONOM_FAKE_STATE=$W/devices.jsonautonom() { python3 $S/scripts/autonom.py --adb $S/tests/fakes/fake_adb.py --simctl $S/tests/fakes/fake_simctl.py --idb $S/tests/fakes/fake_idb.py "$@"; }cat > $AUTONOM_FAKE_STATE <<JSON{"devices": [["emulator-5554", "device", "product:sdk_gphone64_arm64"], ["R5CT20ABCDE", "device", "product:a54x"]], "simctl_devices": {"devices": {"com.apple.CoreSimulator.SimRuntime.iOS-26-0": [{"udid": "AAAAAAAA-1111-2222-3333-BBBBBBBBBBBB", "name": "iPhone 17 Pro", "state": "Booted", "isAvailable": true}]}}}JSON# See what an emulator can doautonom capabilities --serial emulator-5554# Compare it with an Android phoneautonom capabilities --serial R5CT20ABCDE# Measure the screen with --probeautonom capabilities --udid AAAAAAAA-1111-2222-3333-BBBBBBBBBBBBautonom capabilities --udid AAAAAAAA-1111-2222-3333-BBBBBBBBBBBB --probe# Name the target when several are readyautonom capabilities# Let a flow refuse a target that lacks a capabilityprintf '%s\n' 'schema: autonom.dev/flow/v1' 'appId: com.example.app' 'name: Checkout requests' 'requires:' ' capabilities: [network.capture]' '---' '- launchApp' > $W/checkout.yamlautonom session start --serial emulator-5554 --app-id com.example.app > /dev/nullautonom flow run $W/checkout.yamlautonom session stop > /dev/null

Good to know

  • Ask before a long check, not after it fails. A snapshot costs one command. It tells your agent early that a phone has no simulated location, or that the proxy is not attached yet.

  • A snapshot never changes. Take a new one after you attach the proxy, start a session or switch targets; captured_at says which one you are reading.

  • Use --probe on iOS when the screen matters. Without it, ui.accessibility says whether idb is ready. With it, the answer is what Autonom measured on the current screen. While a session runs, a screen read from the last minute refines the plain snapshot too.

  • An unknown capability name is refused early. flow check rejects a name under requires.capabilities that Autonom does not know, with the list of valid names, so a typo never reaches the device.

  • Setup implies capabilities. A flow whose setup sets a location, permissions, orientation, appearance or network also requires the matching capability, without listing it under requires.

  • Every run keeps its snapshot. A flow run records the snapshot it was checked against, and the agent report repeats it under capabilities.

Reference

The command and the flags that pick its target.

CommandWhat it does
autonom capabilitiesPrints the snapshot of the session's target, or of the only ready target when no session runs.
autonom capabilities --probeReads the current screen once first, so the UI entries describe what was measured.
--serial SERIALPicks an Android emulator or phone by its adb serial.
--udid UDIDPicks an iOS Simulator by its UDID.
--target TARGETPicks a target by id, such as one of the candidates of ambiguous_target.
--platform android|iosLimits the choice to one platform.
--session-id IDBinds the command to one session when several run.

The capabilities a snapshot lists.

CapabilityWhat it covers
ui.accessibilityReading the screen as a UI tree.
ui.inputTaps, swipes, keys and typed text.
screenshots, screen.streamScreenshots, and the live view in the Mobile Canvas.
logsDevice logs.
network.captureRequests through the capture proxy. Available once a session has attached it.
checkpoint.create, checkpoint.restoreSaving and restoring a point in a flow. Degraded on an emulator or Simulator, where restoring replays the flow from its start; unavailable on a phone.
simulator.location, simulator.permissionsSimulated location and permission changes.
simulator.appearance, simulator.text_size, simulator.status_bar, simulator.battery, simulator.clipboard, simulator.biometric, simulator.keyboardThe simulator controls both platforms share, where the platform allows them.
simulator.network, simulator.sms, simulator.callNetwork shaping, text messages and calls: Android emulator only.
simulator.pushPush notifications: iOS Simulator only.

Android and iOS differences

Both providers answer with the same snapshot; the values differ with the platform and the kind of device.

CapabilityAndroidiOS Simulator
Device classsimulator on an emulator, physical on a phonesimulator
UI tree and inputFrom adb; --probe changes nothingFrom idb or XCUITest; --probe measures the screen
CheckpointsPartly: degraded on an emulator, unavailable on a phonePartly: degraded
Network, SMS and callsYes, on an emulatorNo
PushNoYes
Keyboard preferencesNoPartly: applied at boot

Next steps

You can see what a target supports before your agent relies on it.

Search documentation

Type to find a guide.

↑ ↓ to move · Enter to open · Esc to close