imsg-media

Full iMessage multimedia pipeline:

🎙️ Voice memo → text via Silicon Flow ASR (SenseVoiceSmall, cloud, no local model)
🖼️ Image → description/OCR via agent's built-in vision model

Requirements

macOS permissions

Full Disk Access must be granted to the process running OpenClaw
Settings → Privacy & Security → Full Disk Access → add your terminal/app
Without this, imsg cannot read ~/Library/Messages/chat.db and will return permissionDenied

API key (audio only)

Silicon Flow API key — sign up free at https://siliconflow.cn
Long-term use: add to ~/.openclaw/.env: SILICON_FLOW_KEY=sk-...
Quick test / override: pass --api-key sk-... directly to the script
Image analysis does not require this key

CLI dependency

imsg CLI: npm install -g imsg

Trigger conditions

Activate this skill when:

Incoming message text contains the attachment placeholder
User says "语音转文字", "转写", "识别语音", "transcribe"
User says "看图", "识别图片", "读图", "OCR", "截图里写的什么"
User references a photo/audio/file they just sent via iMessage

Decision flow

Attachment detected?
├── Audio (.m4a / .caf / .wav / .mp3)  → transcribe via Silicon Flow ASR
├── Image (.jpg / .png / .heic / .gif) → read with vision model
└── Unknown / not downloaded            → increase --limit or ask user to resend

Workflow

Step 1 — Get the sender identifier

Always read from the message envelope:

[iMessage sender@example.com ...] → use sender@example.com
[SMS +1234567890 ...] → use +1234567890
Never hardcode an address

Step 2 — Fetch the attachment

# Run from the skill directory
cd ~/.openclaw/skills/imsg-voice-transcribe

python3 scripts/imsg_voice_transcribe.py fetch \
  --identifier "sender@example.com" \
  --limit 50

Returns JSON with file, type (audio or image), and metadata.

If nothing found, try --limit 100.

Step 3a — Audio: transcribe

# One-liner (fetch + transcribe)
python3 scripts/imsg_voice_transcribe.py auto \
  --identifier "sender@example.com" \
  --limit 50 --raw

# Or transcribe a specific file
python3 scripts/imsg_voice_transcribe.py transcribe \
  --file /path/to/audio.m4a --raw

# Quick test with explicit API key (no env setup needed)
python3 scripts/imsg_voice_transcribe.py transcribe \
  --file /path/to/audio.m4a --api-key sk-... --raw

Step 3b — Image: analyze

After fetch returns an image path (e.g. {"file": "/path/to/photo.jpg", "type": "image"}):

# Example: fetch image from a sender
python3 scripts/imsg_voice_transcribe.py fetch \
  --identifier "sender@example.com" --type image --limit 50
# → {"file": "/Users/.../Messages/Attachments/photo.jpg", "type": "image", ...}

Then in the agent:

If HEIC/HEIF: convert first → sips -s format png "input.heic" --out "output.png"
Open with the read tool → agent vision model processes it
Respond with: what it is, main subject, any text/OCR, notable details

Default image response format:

What it is: photo / screenshot / document
Main subject: 1–2 sentences
Text (OCR): quote key text, or "无明显文字"
Details: 3–5 bullets
Follow-up: ask if they want OCR / table extraction / comparison / etc.

Supported formats

| Format | Type | Notes | |--------|------|-------| | .m4a | Audio | Standard iMessage voice memo | | .caf | Audio | Older iOS voice memo (AAC in CAF) | | .wav .mp3 | Audio | Other sources | | .jpg .jpeg .png | Image | Standard photos | | .heic .heif | Image | iPhone default — convert to PNG first | | .gif | Image | Animated or static |

Troubleshooting

| Error | Cause | Fix | |-------|-------|-----| | permissionDenied | No Full Disk Access | Grant FDA in System Settings | | SILICON_FLOW_KEY not set | Missing API key | Add to ~/.openclaw/.env | | No attachments found | Low limit or iCloud not synced | Increase --limit; ask user to resend | | Request timed out | Network or large file | Retry; check file < 25MB | | HEIC not displaying | Format not supported by read | Convert with sips first |