Configuration

Everything on this page was read from the source. Where the code does something that looks unintended, the page says what the code does and marks it TODO-VERIFY.

config.sh

Created from config.example.sh by setup.sh, then copied to ~/.claude/screen-to-phone-config.sh. It is gitignored.

Variable	Example default	Used by	What the code does with it
BUDDY	+1XXXXXXXXXX	send, send-image, send-video, ask, listener (and receive/agent via env)	Opens imessage://$BUDDY to pick the thread. imessage-ask.sh uses it as the AppleScript buddy. With the + removed it is also matched against chat.chat_identifier when reading replies
APPLE_ID_EMAIL	you@example.com	ask, listener, receive, agent	If set, the reply readers match chats whose identifier contains this value; if empty they fall back to BUDDY. The phone-number match is always added too
FFMPEG	/opt/homebrew/bin/ffmpeg	imessage-send-video.sh	Path to the encoder used to compress videos over 95 MB. ffprobe is called from PATH, not from this variable
RECORDINGS_DIR	$HOME/Desktop/Screen Recordings	nothing	Not read by any script. Studio Record and the browser agent hard-code ~/Desktop/Screen Recordings
TTS_URL	http://localhost:8000/tts	nothing	Not read by any script. build_production_video.py has its own TTS_URL constant with the same value

How scripts find config

Each script that loads config tries, in order:

1. $(dirname "$0")/../config.sh
2. ~/.claude/screen-to-phone-config.sh

When run from the repo (scripts/…), step 1 is the repo's config.sh. When run from ~/.claude/, step 1 resolves to ~/config.sh in your home directory. TODO-VERIFY: if a file named ~/config.sh exists for some other reason, the installed scripts will source it instead of the installed copy.

imessage-receive.sh and imessage-agent.sh do not load config. Run on their own they only see BUDDY and APPLE_ID_EMAIL if exported in the environment. When both are empty, the LIKE '%%' filter matches every conversation. (imessage-agent.sh still sends through imessage-send.sh, which does load config.)

Script arguments

Script	Arguments	Defaults	Output
imessage-send.sh	"message"	none	none on success
imessage-send-image.sh	/path/to/image	none	Sent: <path>
imessage-send-video.sh	/path/to/video	none	Video: …, optional Compressed to …MB, Sent: …
imessage-receive.sh	[timeout_s] [since_rowid] [sent_text]	300, 0 (= latest row now), empty	The newest message text after since_rowid, or TIMEOUT (exit 1). Polls every 3 s
imessage-ask.sh	"what" ["why"] [timeout_s]	timeout 600	YES, NO, YES_TO_ALL, TIMEOUT, or the raw reply
imessage-agent.sh	"question" [timeout_s]	timeout 600	same as ask
imessage-toggle.sh	none		ON (listener PID …, target …) or OFF

Reply keywords (ask and agent)

The reply is lower-cased and trimmed, then matched exactly:

Returns	Replies that match
YES	yes, y, ok, do it, go, go ahead, yes 👍, 👍, yep, yeah, sure, approved
YES_TO_ALL	yes to all, yta, yes all, approve all, do everything (also creates imessage-yesall)
NO	no, n, stop, dont, don't, cancel, reject, nope, 👎
raw text	anything else

Video compression

imessage-send-video.sh compresses when du -m reports more than MAX_MB=95. Target video bitrate in kbit/s is (95 × 8192) / duration × 0.85, encoded with h264_videotoolbox, AAC audio at 128k, +faststart. Output goes to /tmp/imessage-video-compressed.mp4 (overwritten each time).

State files

All live next to the scripts, so ~/.claude/ after install. The statusline and browser agent hard-code ~/.claude/.

File	Written by	Meaning
imessage-agent-on	toggle	Mobile Mode flag. The listener loops while it exists; the browser agent texts updates only when it exists
imessage-mobile-target	toggle	TARGET_TTY=/dev/ttysNNN for the listener
imessage-listener.pid	toggle	PID of the background listener
imessage-sent-recent	send	epoch<TAB>text of recent sends, trimmed to 50 lines. The listener skips an incoming text that matches one sent in the last 60 s
imessage-ask-active	ask, agent	While present the listener does not forward and moves its pointer past new messages
imessage-startup-lock	your launcher (not in repo)	While present the listener waits and does not advance its pointer, so texts are forwarded after startup
imessage-yesall	ask, agent	Every later ask returns YES without sending. Delete it to turn this off
/tmp/imessage-listener.log	toggle	Listener stdout/stderr

Statusline

scripts/statusline.sh reads Claude Code's statusline JSON on stdin and uses .context_window.used_percentage (needs jq; falls back to 0).

• Bar: 10 cells. Color red at 90% and up, yellow 70 to 89, cyan 40 to 69, green below 40.
• Mobile flag: ON only if ~/.claude/imessage-agent-on and ~/.claude/imessage-listener.pid both exist and that PID is alive.
• Agent tag from CLAUDE_SESSION_LABEL (matched by prefix):

CLAUDE_SESSION_LABEL starts with	Tag
Llama 70B	LLAMA 70B
Gemma 4 Browser	GEMMA BROWSER
Gemma 4	GEMMA CODE
Browser Agent	BROWSER AGENT
Claude Code	CLAUDE CODE
NarrativeClaude	NARRATE
empty	no tag
anything else	the label as-is

The script defines a YESALL_FLAG path but does not display it.

Browser agent environment variables

From browser-agent.py:

Env var	Default	Effect
MLX_URL	http://localhost:4000	Base URL; the agent POSTs to $MLX_URL/v1/messages with only a Content-Type header (no API key)
MLX_MODEL_NAME	claude-sonnet-4-6	Sent as model in each request. The server decides what this means
CDP_URL	http://127.0.0.1:9222	Chrome DevTools endpoint. The agent attaches to the first page target
MAX_STEPS	30	Tool calls per task before giving up

Command line: python browser-agent.py "task" runs one task and exits. With no arguments it loops on a What should I do? prompt; quit, exit or q leaves.

Fixed values in the code:

Setting	Value
Request	max_tokens 1024, temperature 0.3, 120 s timeout
History	first message plus last 8 kept once over 10
Tool result	cut at 4000 chars
shell	/bin/bash -lc in $HOME, default timeout 60 s, output cut at 3500 chars
read_file	first 8000 bytes; directories list up to 200 entries
Downloads	~/Downloads/gemma-agent/
Media send	refuses files under 5000 bytes, encrypted-tbn0..3.gstatic.com image URLs, and repeats in the same process
Phone summary	trimmed to 500 chars, sent only when Mobile Mode is on

Tools the model can call: navigate, snapshot, click, type_text, scroll, js, shell, read_file, write_file, screenshot, fullscreen_shot, send_image, send_video, record_start, record_stop, record_status, done. The dispatcher also handles a comment tool that drafts (does not submit) a comment in a page's comment widget. It is not listed in the system prompt. A generate_comment() helper exists but nothing calls it.

The shell tool runs anything

shell and write_file run with your user's permissions and no confirmation step. Whatever the model returns gets executed.

Studio Record

From studio-record/studio_record.py. There are no flags or env vars; these are constants at the top of the file.

Constant	Value
FFMPEG	/opt/homebrew/bin/ffmpeg
OUT_DIR	~/Desktop/Screen Recordings
BG_DIR	~/Desktop/Screen Recordings/backgrounds (.jpg .jpeg .png .webp .bmp)
MODEL_PATH	~/Desktop/Screen Recordings/selfie_segmenter.tflite

HTTP API on 127.0.0.1:17494:

Method + path	Query	Returns
POST /start	mode=screen (default), face, screen_face; unknown values become screen	{"status":"recording","outpath":…,"mode":…}
POST /stop		{"status":"stopped","outpath":…}; waits up to 30 s for the file to appear
GET /status		{"recording":bool,"outpath":…,"mode":…,"duration":seconds}

Output names: screen-YYYYmmdd-HHMMSS.mp4, facecam-…, facecam-bg-…, screen-facecam-…, screen-facecam-bg-…. Video is 30 fps h264_videotoolbox at 10 Mbit/s; audio is 48 kHz mono 16-bit from sounddevice. Microphones whose names contain usb, rode, yeti or shure are preferred, then a MacBook mic, then the system default.

build_production_video.py

A template, not a tool. The paths, voice-over text, keep-segments and subtitle lines are hard-coded constants at the top and in the body. It has no arguments. Edit INPUT, LOGO, OUTPUT, TTS_URL, the segs list and subtitle_data before running. It expects macOS system fonts in /System/Library/Fonts/ and a TTS server that accepts POST with a text form field.