diff --git a/api/device-sessions.md b/api/device-sessions.md
index 7c68e0c..51d5929 100644
--- a/api/device-sessions.md
+++ b/api/device-sessions.md
@@ -20,7 +20,7 @@ A session is named by its `id` (for example `brave-teal-otter`) in every path. E
 |---|---|
 | `POST /device/sessions/:id/install` | Install a build: `{ "binaryUploadId": "..." }`, or `{ "url": "https://..." }` for an `.apk` or a zipped simulator `.app` at a link (up to 500 MB). It installs in the background. A command sent meanwhile waits up to 45 s for it, then answers `LIVE_NOT_READY` |
 | `POST /device/sessions/:id/launch` | Launch an app: `{ "appId"?: "...", "clearState"?: true }`. Without `appId`, the installed build |
-| `POST /device/sessions/:id/actions` | One action: `tap` (`x`/`y` percent, `text` or `id`), `type` (`text`, and `sensitive: true` to keep it secret), `swipe` (`direction`: `up`, `down`, `left` or `right`, or `from`/`to` points in percent), `press_key` (`key`: `back`, `home`, `enter`, `backspace`, `tab`, `volume_up`, `volume_down`, `lock` or `power`) |
+| `POST /device/sessions/:id/actions` | One action: `tap` (`x`/`y` percent, `text` or `id`), `type` (`text`, and `sensitive: true` to keep it secret), `swipe` (`direction`: `up`, `down`, `left` or `right`, or `from`/`to` points in percent), `press_key` (`key`: `back`, `home`, `enter`, `backspace`, `tab`, `volume_up`, `volume_down`, `lock` or `power`). Performed on the device directly and answered once the screen has settled, usually well under a second; `text` and `id` are matched whole, ignoring case. `mode: "maestro"` runs it as a Maestro flow instead. A `tap` on something that is not on screen waits about 17 s before it fails |
 | `POST /device/sessions/:id/run` | Run Maestro: `{ "yaml": "..." }` for one flow, or `{ "flowUploadId": "...", "flow"?: "..." }` for a [flow bundle](#flow-bundles). `env` passes [secrets and variables](#secrets). Waits up to 110 s, or `waitSeconds` (0 to 300), then answers `running` with a `command_id`. A failed run's result includes `screen_text`, the app's visible text after the failure |
 | `GET /device/sessions/:id/commands/:commandId` | The status and result of a `run` that answered `running`. `?waitSeconds=` (0 to 300) waits for it to finish instead of answering at once |
 | `GET /device/sessions/:id/inspect` | The screen as compact JSON. The on-screen keyboard and the System UI are left out unless you add `?all=true` |
diff --git a/cli/dcd-device.md b/cli/dcd-device.md
index 94539a4..ae88ad8 100644
--- a/cli/dcd-device.md
+++ b/cli/dcd-device.md
@@ -16,6 +16,8 @@ A typical session:
 dcd device start --platform ios --app ./build/MyApp.app --wait
 dcd device launch
 dcd device inspect
+dcd device tap --text "Log in"
+dcd device type --text "ann@example.com"
 dcd device run flows/login.yaml
 dcd device screenshot -o after-login.jpg
 dcd device stop
@@ -95,6 +97,31 @@ The screen as compact JSON: each element with its text, id, accessibility label
 
 Save a screenshot. `-o <file>` sets where (default `./<session>-<time>.jpg`). With `--json`, `data` gives the file's path.
 
+## `dcd device tap [id]`
+
+Tap one thing on the screen, by exactly one target. Answers once the screen has settled, usually well under a second. A target that is not on screen waits about 17 seconds before it fails, so `dcd device inspect` first. A failed tap exits `2`.
+
+| Flag | Description |
+|------|-------------|
+| `--text <text>` | The element's text, label or hint, matched whole and ignoring case (not a regex) |
+| `--rid <id>` | The element's resource id, `rid` in `dcd device inspect` |
+| `--x <n> --y <n>` | A point, as percentages of the screen (0 to 100) |
+| `--maestro` | Run the tap as a Maestro flow instead of directly on the device |
+
+## `dcd device type [id] --text <text>`
+
+Type into the focused field (tap it first). `--sensitive` keeps the text secret: it is stored encrypted, never shown, and masked in the output. `--maestro` as above.
+
+## `dcd device swipe [id] <up|down|left|right>`
+
+Swipe across most of the screen in a direction (`up` scrolls the content up, down the page), or from one point to another with `--from x,y --to x,y`, in percentages. `--maestro` as above.
+
+## `dcd device press-key [id] <key>`
+
+Press a key: `back`, `home`, `enter`, `backspace`, `tab`, `volume_up`, `volume_down`, `lock` or `power`. `enter` submits the focused field; `back` closes a keyboard or goes back on Android. `--maestro` as above.
+
+Each of these sends one [typed action](../api/device-sessions.md). For several steps, an assertion, scrolling until something is visible, or launching an app, use `run`.
+
 ## `dcd device run [id] <flow.yaml|directory>`
 
 Run Maestro on the device and wait for the result. A failed flow exits `2`. Maestro's output is printed either way, and after a failure, so is the app's visible text.
diff --git a/mcp/agent-devices-skill.md b/mcp/agent-devices-skill.md
index 465ff94..57636ca 100644
--- a/mcp/agent-devices-skill.md
+++ b/mcp/agent-devices-skill.md
@@ -14,7 +14,7 @@ Use the DeviceCloud MCP tools to check app changes on a real iOS simulator or An
 1. Call `cheat_sheet` once for Maestro syntax.
 2. `start_device` with the platform (`list_devices` shows the models). Give the user the session's `monitor_url`, where they can watch. A queued session is normal: it starts on its own. Until it is ready, call `get_device_status` with `wait: true`.
 3. `install_app`, then `get_device_status` with `wait: true` until the install is done, then `launch_app`.
-4. Loop: `inspect_screen` to see what is on screen, then `run` with Maestro commands to act, then `inspect_screen` again. Target elements by their text or id from `inspect_screen`, not by coordinates. Use `take_screenshot` when the layout matters.
+4. Loop: `inspect_screen` to see what is on screen, then act: `tap_on`, `input_text`, `swipe` or `press_key` for one step (the fastest path), or `run` with Maestro commands for several steps or an assertion, then `inspect_screen` again. Target elements by their text or id from `inspect_screen`, not by coordinates. Use `take_screenshot` when the layout matters.
 5. If a step needs a person (a 2FA code, a CAPTCHA, a payment sheet), call `request_human` with what they should do, tell the user, and poll `get_human_status` until it is handed back.
 6. Always `stop_device` when you are done, even after a failure. A session bills by the minute.
 
@@ -37,7 +37,7 @@ description: Check app changes on a real cloud iOS simulator or Android emulator
 1. Call `cheat_sheet` once for Maestro syntax.
 2. `start_device` with the platform, and give the user the session's `monitor_url`. A queued session starts on its own: call `get_device_status` with `wait: true` until it is ready.
 3. `install_app` (with `app_path` for a local build when using `dcd mcp`), wait with `get_device_status` until it is installed, then `launch_app`.
-4. Loop: `inspect_screen`, then `run` with Maestro commands, then `inspect_screen` again. Target elements by text or id, not coordinates.
+4. Loop: `inspect_screen`, then `tap_on`, `input_text`, `swipe` or `press_key` for one step, or `run` with Maestro commands for several, then `inspect_screen` again. Target elements by text or id, not coordinates.
 5. If a person is needed (2FA, CAPTCHA, payment), call `request_human`, tell the user, and poll `get_human_status` until it is handed back.
 6. Always `stop_device` when done, even after a failure.
 
diff --git a/mcp/agent-devices-tools.md b/mcp/agent-devices-tools.md
index 47cfaaf..53fd03d 100644
--- a/mcp/agent-devices-tools.md
+++ b/mcp/agent-devices-tools.md
@@ -116,9 +116,51 @@ Read-only.
 |-------|------|----------|-------------|
 | `session_id` | string | No | The session id from start_device, e.g. "brave-teal-otter". Optional when your organization has exactly one open session. |
 
+## `tap_on`
+
+Tap one thing on the screen: an element by its text (txt or a11y from inspect_screen, matched whole and ignoring case, not as a regex), by its id (rid), or a point as x and y percentages of the screen. Give exactly one target. The fastest way to act: it answers once the screen has settled, usually well under a second. A target that is not on screen waits about 17 seconds before it fails, so inspect_screen first. For several steps or an assertion, use run.
+
+| Input | Type | Required | Description |
+|-------|------|----------|-------------|
+| `session_id` | string | No | The session id from start_device, e.g. "brave-teal-otter". Optional when your organization has exactly one open session. |
+| `text` | string | No | The element's text, label or hint, copied from inspect_screen. |
+| `id` | string | No | The element's resource id (rid in inspect_screen). |
+| `x` | number, 0 to 100 | No | x as a percentage of the screen, 0 to 100. |
+| `y` | number, 0 to 100 | No | y as a percentage of the screen, 0 to 100. |
+
+## `input_text`
+
+Type into the focused field (tap_on the field first). Answers once the screen has settled. With sensitive: true the text is stored encrypted, never shown and masked in the output: use it for passwords and codes.
+
+| Input | Type | Required | Description |
+|-------|------|----------|-------------|
+| `session_id` | string | No | The session id from start_device, e.g. "brave-teal-otter". Optional when your organization has exactly one open session. |
+| `text` | string | Yes | What to type, as the user would. |
+| `sensitive` | boolean | No | The text is a secret: a password, a code. Default false. |
+
+## `swipe`
+
+Swipe across the screen: a direction (up scrolls the content up, i.e. down the page), or from one point to another as x and y percentages. Give either direction, or from and to. Answers once the screen has settled.
+
+| Input | Type | Required | Description |
+|-------|------|----------|-------------|
+| `session_id` | string | No | The session id from start_device, e.g. "brave-teal-otter". Optional when your organization has exactly one open session. |
+| `direction` | `up`, `down`, `left` or `right` | No | A swipe across most of the screen. |
+| `from` | object `{ x, y }` | No | Where the swipe starts, in percentages of the screen. |
+| `to` | object `{ x, y }` | No | Where it ends. |
+
+## `press_key`
+
+Press a device key: back (Android), home, enter, backspace, tab, volume_up, volume_down, lock, power. enter submits the focused field; back closes a keyboard or goes back. Answers once the screen has settled.
+
+| Input | Type | Required | Description |
+|-------|------|----------|-------------|
+| `session_id` | string | No | The session id from start_device, e.g. "brave-teal-otter". Optional when your organization has exactly one open session. |
+| `key` | `back`, `home`, `enter`, `backspace`, `tab`, `volume_up`, `volume_down`, `lock` or `power` | Yes | Which key. |
+
 ## `run`
 
-Run Maestro on the device: `yaml` is one flow (a header naming the app, `---`, then commands), run whole. Or `flow_upload_id` runs an uploaded zip of flows (`flow` names the entry file; omitted, all of them run). Waits up to 110s (wait_seconds changes it); a longer flow answers status "running" with a command_id: call get_command_status with wait: true. A failed run's result has screen_text, the app's visible text after the failure, and failed_at: the file, line and command that failed. See cheat_sheet for syntax.
+Run Maestro on the device: `yaml` is one flow (a header naming the app, `---`, then commands), run whole. Or `flow_upload_id` runs an uploaded zip of flows (`flow` names the entry file; omitted, all of them run). For a single tap, swipe, key or typed text prefer tap_on, swipe, press_key and input_text, which are faster; use run for several steps, assertions, scrolling until something is visible, or launching an app. Waits up to 110s (wait_seconds changes it); a longer flow answers status "running" with a command_id: call get_command_status with wait: true. A failed run's result has screen_text, the app's visible text after the failure, and failed_at: the file, line and command that failed. See cheat_sheet for syntax.
 
 | Input | Type | Required | Description |
 |-------|------|----------|-------------|
diff --git a/mcp/agent-devices.md b/mcp/agent-devices.md
index 9b02f20..b183bc4 100644
--- a/mcp/agent-devices.md
+++ b/mcp/agent-devices.md
@@ -108,7 +108,7 @@ Every client you sign in shows up under **Settings → Account → Connected app
 1. `list_devices` shows the devices it can start, and any sessions already open.
 2. `start_device` leases a device and waits up to a minute for it. If the fleet is busy, the session is queued, which isn't an error: it starts on its own, and the agent waits for it with `get_device_status`.
 3. `install_app` installs a build you uploaded, or one at an HTTPS link. `launch_app` starts it.
-4. The agent loops: `inspect_screen` to see what's on screen, then `run` to act with Maestro commands, then `inspect_screen` again. `take_screenshot` shows the screen as an image, and `get_logs` reads the app's device log.
+4. The agent loops: `inspect_screen` to see what's on screen, then `tap_on`, `input_text`, `swipe` or `press_key` for one step, or `run` for several Maestro commands, then `inspect_screen` again. `take_screenshot` shows the screen as an image, and `get_logs` reads the app's device log.
 5. `stop_device` releases the device when it's done.
 
 Every response carries a short **session** block: the session's state, how long it has left, its cost so far, its page in the console, and plain-English warnings, such as an upcoming idle timeout or the app having crashed. The agent never needs a separate call to stay oriented. Agents are told to give you the session's console link as soon as they have it, so you can watch.
diff --git a/public/openapi.yaml b/public/openapi.yaml
index 19f48cc..331a14f 100644
--- a/public/openapi.yaml
+++ b/public/openapi.yaml
@@ -1422,11 +1422,13 @@ paths:
       operationId: performAction
       summary: Perform one action
       description: >-
-        Sends one action to the device: `tap`, `type`, `swipe` or `press_key`. Each becomes one Maestro
-        command. The request waits for it to finish, up to 120 s, after waiting up to 45 s for a device that
-        is still starting or installing. An action that fails on the device still answers 201, with
-        `result.success` false. A `type` action with `sensitive: true` sends its text as a secret: it never
-        appears in the timeline and is masked in `inspect`.
+        Sends one action to the device: `tap`, `type`, `swipe` or `press_key`. The device performs it
+        directly and the request answers once the screen has settled, usually well under a second (up to
+        120 s, after waiting up to 45 s for a device that is still starting or installing). `mode: "maestro"`
+        runs the action as a one-command Maestro flow instead. A `tap` on something that is not on screen
+        waits about 17 s before it fails, so `inspect` first. An action that fails on the device still answers
+        201, with `result.success` false. A `type` action with `sensitive: true` sends its text as a secret: it
+        never appears in the timeline and is masked in `inspect`.
       parameters:
         - name: x-dcd-source
           in: header
@@ -4073,12 +4075,15 @@ components:
       type: object
       description: >-
         Tap one target: a point (`x` and `y`), an element's `text`, or its `id`. Exactly one. Text and ids are
-        matched literally, not as regular expressions.
+        matched whole and ignoring case, literally, not as regular expressions. A target that is not on
+        screen waits about 17 s before the tap fails.
       required: [type]
       properties:
         type:
           type: string
           const: tap
+        mode:
+          $ref: '#/components/schemas/ActionMode'
         x:
           type: number
           minimum: 0
@@ -4112,6 +4117,8 @@ components:
         type:
           type: string
           const: type
+        mode:
+          $ref: '#/components/schemas/ActionMode'
         text:
           type: string
           minLength: 1
@@ -4131,6 +4138,8 @@ components:
         type:
           type: string
           const: swipe
+        mode:
+          $ref: '#/components/schemas/ActionMode'
         direction:
           type: string
           enum: [up, down, left, right]
@@ -4150,10 +4159,19 @@ components:
         type:
           type: string
           const: press_key
+        mode:
+          $ref: '#/components/schemas/ActionMode'
         key:
           type: string
           enum: [back, home, enter, backspace, tab, volume_up, volume_down, lock, power]
 
+    ActionMode:
+      type: string
+      enum: [maestro]
+      description: >-
+        `maestro` runs the action as a one-command Maestro flow instead of directly on the device: slower,
+        with Maestro's own waits. Leave it out otherwise.
+
     ScreenPoint:
       type: object
       description: A point on the screen, in percent. Each value is rounded to a whole number.