Imsg Media

Fetch iMessage/Messages.app attachments (voice memos and images) and process them — transcribe audio via Silicon Flow ASR (SenseVoiceSmall), and analyze imag...

Security Scan

The skill's code, instructions, and required environment variable line up with its stated purpose (fetch iMessage attachments and transcribe images/audio), but it requires Full Disk Access and will upload audio to a third‑party ASR service — review those privacy implications before installing.

Purpose & Capability

Name/description match what the files do: the script uses the 'imsg' CLI to locate iMessage attachments and sends audio to Silicon Flow for transcription. Required binary (imsg) and env var (SILICON_FLOW_KEY) are expected for this functionality. Minor inconsistency: SKILL.md asks to cd into '~/.openclaw/skills/imsg-voice-transcribe' while the registry slug is 'imsg-media' (directory/name mismatch), but this is an operational nit, not a functional mismatch.

Instruction Scope

Runtime instructions and the script operate within the declared scope: they read Messages attachments via the imsg CLI, classify files as audio/image, upload audio to the declared Silicon Flow API, and instruct the agent to use its vision model for images. The skill explicitly requires Full Disk Access so it can read ~/Library/Messages attachments; that is necessary but also grants broad access to message contents (see guidance). The script only reads the specified env file (~/.openclaw/.env) for the API key and does not attempt to read unrelated system config.

Install Mechanism

This is instruction-only with one included script; there is no automated install or remote download. The only external install the SKILL.md recommends is 'npm install -g imsg' for the imsg CLI, which is reasonable and expected.

Credentials

The single required credential (SILICON_FLOW_KEY) is appropriate for the declared cloud transcription service, but the skill requires Full Disk Access to read Messages data — a high privilege that exposes all message attachments and metadata. Additionally audio files are uploaded to https://api.siliconflow.cn, so granting the API key and FDA has privacy/egress implications. The script only reads the stated ~/.openclaw/.env file for the key and supports a one-off --api-key override (good), but users should confirm they trust the external ASR provider before storing a long‑lived key.

Persistence & Privilege

The skill does not request always:true, does not modify other skills' configs, and has no install-time persistent service. It suggests adding the API key to ~/.openclaw/.env for convenience, which is normal but increases persistent credential exposure if used.

Assessment

This skill does what it says — it locates iMessage attachments (requires the imsg CLI) and uploads audio files to the Silicon Flow ASR endpoint for transcription. Before installing: (1) understand that you must grant Full Disk Access to the process running OpenClaw/your terminal — that allows reading all Messages data and attachments; (2) review Silicon Flow's privacy/security policy because audio will be sent to api.siliconflow.cn; (3) prefer using a one-time --api-key override instead of putting a long-lived key in ~/.openclaw/.env if you want less persistence; (4) verify the imsg CLI you install is from a trusted source; (5) note the small filename/slug mismatch in the README (cd path) and check the skill directory name before running. If you are uncomfortable granting FDA or sending audio to an external service, do not install or run this skill.

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 [e-mail] ...] → use [e-mail]
• [SMS [phone] ...] → use [phone]
• 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 "[e-mail]" \
  --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 "[e-mail]" \
  --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 "[e-mail]" --type image --limit 50
# → {"file": "/Users/.../Messages/Attachments/photo.jpg", "type": "image", ...}

Then in the agent:

1. If HEIC/HEIF: convert first → sips -s format png "input.heic" --out "output.png"
2. Open with the read tool → agent vision model processes it
3. 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