Vision
- Sensor
- OV5640-AF, MIPI CSI
- Resolution
- 5 MP
- Default capture 1
- 1080p
- Capture latency
- <500 ms
- Autofocus
- VCM, 4.5 mm
- ISP
- RV1106 native (debayer, AWB, AE)
- Frame attestation
- Secure element, ECDSA
The full numbers
Provisional. Final values land at design freeze. The datasheet is the truth — if a marketing line and the spec disagree, the spec wins.
Harness presents itself to a Claude Code, Roo, or any MCP host as a flat surface of tools. No SDK to install. No state to manage. The agent calls; the device responds.
| Tool | Returns | Note |
|---|---|---|
| 01 capture | image (JPEG) | single frame, < 500 ms |
| 02 listen | transcript or WAV | VAD on device, STT in cloud |
| 03 say | duration | TTS playback |
| 04 ask | transcript | say + ding + listen |
| 05 set_yaw | angle | absolute, ±135° |
| 06 set_yaw_mode | mode | passive · cooperative · locked |
| 07 set_led | void | color + pattern |
| 08 wait_for_sound | detection | blocking; agent sleeps until trigger |
| 09 wait_for_change | frame + diff | blocking; visual change detection |
| 10 start_monitoring | status | ring buffer, 1–60 s |
| 11 get_state | object | full device snapshot |
| 12 get_telemetry | metrics | latency, errors, motor health |
Subset shown. The full tool list locks at design freeze.