Camera — Physical Photos & Video¶

Drive the actual camera. Not screenshots — real photos from the lens.

adb(action="camera_photo", facing="back")
adb(action="camera_photo", facing="front")
adb(action="camera_video", facing="back", duration_sec=10)

Returns the captured media as a Converse API image/video block, so the agent can immediately see what it just captured.

How It Works¶

sequenceDiagram
    participant Agent
    participant Tool as adb tool
    participant Cam as GoogleCamera
    participant FS as /sdcard/DCIM

    Agent->>Tool: camera_photo, facing="front"
    Tool->>Tool: note latest DCIM file
    Tool->>Cam: am start STILL_IMAGE_CAMERA
    Note over Cam: viewfinder up
    Tool->>Cam: smart_tap "Switch to front camera"
    Tool->>Cam: tap shutter_button (by resource-id)
    Cam->>FS: write IMG_xxx.jpg
    Tool->>FS: poll for new file
    FS-->>Tool: /sdcard/DCIM/Camera/IMG_xxx.jpg
    Tool->>Tool: adb pull → /tmp/IMG_xxx.jpg
    Tool-->>Agent: image block (JPEG bytes)

Parameters¶

Param	Default	Notes
`facing`	`"back"`	`"back"` or `"front"`
`output_path`	`/tmp/<original>.jpg`	Where to pull the file
`auto_pull`	`True`	Pull to host after capture
`include_image`	`True`	Embed image block in response
`return_base64`	`False`	Also include base64
`timeout_sec`	`15`	Max wait for file to appear (Night Sight needs ~5s)

Strategy: Why GoogleCamera?¶

Most phones have multiple camera apps. We target GoogleCamera specifically because:

Its shutter button has a stable resource-id: com.google.android.GoogleCamera:id/shutter_button
Google doesn't rename it across versions
Available on all Pixel / most Android-12+ devices

For other camera apps, adapt via resource_id parameter.

Night Sight / Portrait / Other Modes¶

camera_photo currently uses whatever mode GoogleCamera was last in. To force a mode, smart_tap it first:

adb(action="launch", package="com.google.android.GoogleCamera")
adb(action="smart_tap", text="Night Sight")
adb(action="camera_photo", facing="back")  # captures in Night Sight

Video¶

result = adb(action="camera_video", duration_sec=15, facing="back")
# result["path"] → /tmp/VID_xxx.mp4

For longer-form recording, use screen_record instead (records the screen, not camera).

Agent Recipes¶

"Take a selfie and tell me how I look"¶

agent("take a front-facing photo and describe what you see")
# 1. camera_photo facing=front
# 2. vision model sees the photo
# 3. agent describes it

"Document what's in front of me"¶

agent("take a back-camera photo and identify any objects")

"Is my desk tidy?"¶

agent("take a photo and rate my desk tidiness 1-10")

Pitfalls¶

Problem	Fix
Photo never appears (timeout)	Night Sight in low light can take 5-8s — increase `timeout_sec`
Wrong camera faces you	Pass `facing="front"` / `"back"` explicitly
Shutter never fires	Camera UI still animating — we retry 3× by default
Photo ends up blurry	Phone moved during capture — use a stand
Permission denied reading DCIM	Verify `adb shell ls /sdcard/DCIM/Camera` works

What's Next¶

Vision — difference between screenshots and camera
Sensors — ambient light, gyro (auto-select mode)
Smart Tap — how we find the shutter button