imessage-rich-search
Enables full-text search of macOS iMessages including link preview metadata. Works as an MCP server for Claude Desktop to search your messages locally.
README
imessage-rich-search
Full-text search across macOS iMessages — including the link preview metadata (titles, summaries, site names) that Messages.app indexes but the raw
chat.dbtext column never exposes.
The problem this solves
When you paste a URL into iMessage, macOS fetches a rich preview — title, summary, site name, hero image — and stores that metadata in chat.db as an NSKeyedArchiver blob in message.payload_data. Messages.app's search bar reads it. The raw chat.db text column does not.
So if a friend sent https://x.com/foo/status/123 and the preview card said "Obsidian + Claude Code is the most underrated productivity stack" — searching for "obsidian" in any tool that only reads text returns zero results. Messages.app finds it. This tool finds it. They search the same surface.
What it is — and isn't
Is: A read-only, local search over your chat.db. ~200 lines of stdlib Python. Zero runtime dependencies. CLI + optional MCP server for Claude Desktop.
Isn't: A Messages replacement (no UI, no send/edit/delete). A way to access anyone else's messages. An iCloud sync tool — searches only what's locally on this Mac. An OCR / image / audio / sticker reader. A bypass for Full Disk Access — you must grant it explicitly.
Requirements
| OS | macOS 11 Big Sur or newer (tested through macOS 26) |
| Architecture | Apple Silicon (arm64) or Intel (x86_64) |
| Python | Apple's system /usr/bin/python3 (3.9, ships with Xcode Command Line Tools) — required, see Why Apple's Python |
| Permissions | One specific Full Disk Access grant — see Step 2 |
| Disk | ~30 KB code + ~15 MB venv |
If you don't have Command Line Tools, run xcode-select --install first.
Install
Step 1 — Run the installer (one command)
curl -fsSL https://raw.githubusercontent.com/cannavis/imessage-rich-search/main/install.sh | bash
What this does, in plain English:
- Verifies you're on macOS with Command Line Tools.
- Creates a virtual environment at
~/.local/share/imessage-rich-searchagainst Apple's system Python 3.9. - Pulls and installs this package from GitHub into that venv.
- Symlinks three commands into
~/.local/bin:imessage-rich-search,imrs(alias),imessage-rich-search-mcp.
If you'd rather not pipe curl into bash, do it manually:
/usr/bin/python3 -m venv ~/.local/share/imessage-rich-search
~/.local/share/imessage-rich-search/bin/pip install --upgrade pip
~/.local/share/imessage-rich-search/bin/pip install "git+https://github.com/cannavis/imessage-rich-search@v0.2.1"
mkdir -p ~/.local/bin
for exe in imessage-rich-search imrs imessage-rich-search-mcp; do
ln -sf ~/.local/share/imessage-rich-search/bin/$exe ~/.local/bin/$exe
done
The CLI works immediately after this:
~/.local/bin/imrs "obsidian" --limit 3
If your shell can't find imrs, add ~/.local/bin to your PATH:
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && exec zsh
Step 2 — Grant Full Disk Access to Apple's Python 3.9
You must add this exact path to FDA, or the MCP server (and any disclaimed call into chat.db) will fail with authorization denied. This step does not affect the CLI when run from your own terminal.
-
Open System Settings → Privacy & Security → Full Disk Access.
-
Click the + button.
-
In the file picker, press ⌘ + Shift + G (Go to Folder).
-
Paste this path exactly:
/Library/Developer/CommandLineTools/Library/Frameworks/Python3.framework/Versions/3.9/bin -
Hit Return — you'll see the contents of that
binfolder. -
Select
python3.9(notpython3, which is a symlink). -
Click Open. It appears in the list as
python3.9. -
Make sure the toggle is ON. Authenticate if prompted.
Step 3 — Wire the MCP server into Claude Desktop
Edit ~/Library/Application Support/Claude/claude_desktop_config.json and merge this in (preserve any existing keys):
{
"mcpServers": {
"imessage-rich-search": {
"command": "/Users/YOUR_USERNAME/.local/bin/imessage-rich-search-mcp"
}
}
}
Replace YOUR_USERNAME with your actual username (whoami will tell you).
Step 4 — Restart Claude Desktop
⌘Q (full quit — not just close-window) and relaunch. New chats will have a search_imessages_rich tool.
Verify it works
CLI:
imrs "obsidian" --limit 3
# 3 match(es) for 'obsidian':
# [2026-04-08T22:56:23+00:00] -> +1XXXXXXXXXX (rowid=...)
# * preview: Claude Code + Obsidian Ultimate Guide (build an AI second brain)
MCP server (manual JSON-RPC handshake — same path Claude Desktop uses):
(printf '%s\n%s\n' \
'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"test","version":"0"}}}' \
'{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"search_imessages_rich","arguments":{"query":"obsidian","limit":1}}}') \
| /Applications/Claude.app/Contents/Helpers/disclaimer ~/.local/bin/imessage-rich-search-mcp
You should see two JSON-RPC responses. If the second one contains "isError": true and "authorization denied", Step 2 (FDA grant) was missed or the toggle is off.
Usage
imrs "obsidian" # search all conversations
imrs "obsidian" --contact "+14073993471" # restrict to one handle
imrs "claude code" --json | jq '.[].preview[0]' # JSON for piping
imrs "obsidian" --limit 20 --db /backup/chat.db # backup file
imessage-rich-search --help # full options
Output legend: -> sent · <- received · * preview hit · rowid cross-references back into chat.db.
Claude Desktop (MCP server)
Once Steps 1–4 above are done, Claude can call this tool directly:
search_imessages_rich(query, contact?, limit?)
query— case-insensitive substring (required)contact— optional handle filter, e.g."+14073993471"or"name@example.com"limit— max matches, default 50
Returns newest-first matches with body, decoded preview metadata, handle, date, and rowid.
Why Apple's Python
This isn't arbitrary — it's the only thing that works inside Claude Desktop, and there's a clear reason:
Claude Desktop spawns MCP servers through /Applications/Claude.app/Contents/Helpers/disclaimer, a tiny wrapper that calls responsibility_spawnattrs_setdisclaim(). This is an Apple API that deliberately breaks the TCC responsibility chain so the MCP server is treated as its own responsible process. The intent: prevent third-party MCP servers from silently inheriting Claude.app's broad permissions.
Consequence: the spawned binary needs its own Full Disk Access grant — Claude.app's grant doesn't propagate. macOS resolves the venv's python3 to its canonical Apple binary at /Library/Developer/CommandLineTools/.../python3.9, and that's the path TCC checks. Hence Step 2.
If you install with Homebrew Python or pyenv instead, the canonical path resolves somewhere else and TCC blocks the request. Apple's CLT Python is the path of least resistance.
References:
- Qt: The Curious Case of the Responsible Process
- Michael Tsai's notes on
responsibility_spawnattrs_setdisclaim
How it works
chat.db (SQLite, opened with mode=ro)
└─ message
├─ text ← raw text (what basic tools see)
├─ payload_data BLOB ← NSKeyedArchiver bplist of LPLinkMetadata
│ (title, summary, site, image refs)
└─ balloon_bundle_id ← e.g. com.apple.messages.URLBalloonProvider
For every row matching the contact filter:
1. Read text + payload_data
2. plistlib.loads(payload_data) → walk $objects → collect strings
3. haystack = (text + '\n'.join(preview_strings)).lower()
4. Match if query.lower() in haystack
Walking strings out of $objects avoids needing ccl_bplist, pyobjc, or full NSKeyedUnarchiver — for full-text search the leaf strings are all that matter.
Privacy & security
- Read-only. Opens
chat.dbwith SQLite URI flagmode=ro. - Local-only. No network calls. (
grep -r 'urllib\|requests\|http\|socket' src/returns nothing.) - No telemetry.
- Zero runtime dependencies — nothing to be supply-chain-attacked through.
- The FDA grant from Step 2 applies to Apple's system Python 3.9 system-wide. If you have
Terminal,bash, orVisual Studio Codealready in FDA, this isn't expanding your attack surface — those can already trivially shell out to/usr/bin/python3.
See SECURITY.md for vulnerability reporting.
Limitations
- Substring match only. No FTS5, regex, or boolean operators. (Roadmap.)
- Local DB only. If a message lives only in iCloud and isn't synced to this Mac's
chat.db, this tool won't see it. - Preview metadata depends on Messages.app having fetched it. If a link card never loaded (offline send, expired URL), there's no
payload_datato search. - String extraction is lossy by design. Image MIME types, dimension tuples like
{0, 0}, and profile-image URLs may appear in raw output. They don't affect search hits.
Troubleshooting
| Symptom | Fix |
|---|---|
imrs: command not found |
Add ~/.local/bin to PATH: echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && exec zsh |
unable to open database file from CLI in your terminal |
Add your terminal app to FDA: System Settings → Privacy & Security → Full Disk Access. ⌘Q + relaunch the terminal. |
authorization denied from Claude Desktop's MCP call (CLI works fine) |
Step 2 is missing or toggled off. Verify python3.9 is in FDA list with toggle ON. ⌘Q + relaunch Claude Desktop. |
| Claude Desktop doesn't see the tool at all | Validate JSON: python3 -c "import json; json.load(open('$HOME/Library/Application Support/Claude/claude_desktop_config.json'))". Then ⌘Q + relaunch. |
| Returns 0 matches but Messages.app finds them | Wrong --contact format. Drop --contact to confirm. Phone numbers must be E.164: +14155551212. |
| Apple's Python 3.9 isn't at the expected path | xcode-select -p to verify CLT is installed. If installed but path differs (e.g., full Xcode), find it: xcrun --find python3 |
| Search slow on huge DBs | Linear scan + bplist parse per row. ~100k message DBs take a few seconds. FTS5 index is roadmap. |
Verify the TCC grant directly
If unsure whether Step 2 took effect, watch the system log while making a request:
log stream --predicate 'process == "tccd"' --info | grep -E "python3\.9|chat\.db|SystemPolicyAllFiles"
You want to see Auth Right: Allowed (System Set) for the python3.9 binary. Denied (Service Policy) means the grant isn't in place.
Contributing
See CONTRIBUTING.md. Issues and PRs welcome. Don't paste real chat content into public issues — redact phone numbers, names, and message text first.
Changelog
See CHANGELOG.md.
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。