Privacy Monitor AI connector
A local, metadata-and-insight MCP server for Claude/Codex-compatible stdio clients.
Read tools: activity_summary, recent_activity, contacted_domains, and list_insights. publish_insight saves requested AI analysis locally, with real event IDs, a review window, confidence, limitations, and recommended next steps.
No arbitrary SQL, shell commands, monitoring-settings changes, or automatic uploads. Only publish_insight writes, and only to the saved-review table when requested.
Time windows are bounded to seven days and results are capped. Clipboard contents,
raw file paths, complete URLs, query strings, and credential details are omitted.
When an AI client calls tools, returned metadata enters that client's conversation
and is subject to its data handling; the connector cannot promise it stays on-device.
On the iPhone
The Sileo collector package installs this command:
/var/jb/usr/libexec/privacy-monitor/launch-mcp
For development before upgrading the collector, use the absolute path to
privacy-monitor-ai/server/launch in this checkout. It chooses an installed Python
and serves JSON-RPC on stdin/stdout; it does not open a TCP port.
Claude Code registration (supported CLI):
claude mcp add --transport stdio privacy-monitor -- /var/jb/bin/sh /var/jb/usr/libexec/privacy-monitor/launch-mcp
Codex registration:
codex mcp add privacy-monitor -- /var/jb/bin/sh /var/jb/usr/libexec/privacy-monitor/launch-mcp
A new client session may be needed to load tools. Registration and actual client availability must be verified separately; protocol tests alone do not prove a chat has loaded the connector.
From a computer
Use an SSH stdio command to run the installed launcher on the phone. Set up your own key-based SSH access first. Example MCP client configuration:
{"mcpServers":{"privacy-monitor":{"command":"ssh","args":["-T","-o","BatchMode=yes","mobile@YOUR_IPHONE_IP","/var/jb/bin/sh","/var/jb/usr/libexec/privacy-monitor/launch-mcp"]}}}
Replace the address; do not embed a root password. SSH needs access to the phone, while database reads run as mobile. Do not expose the local HTTP dashboard directly to the internet. A browser-only/mobile-cloud AI client cannot launch this stdio server without a separate supported bridge; none is provisioned here.
The portable local plugin is this directory (plugin.json, mcp.json, skills/, server/, and assets/). It runs on the jailbroken device, not on arbitrary desktop systems without SSH. No cloud plugin publication is included.
Verification
Four isolated tests cover protocol initialization/discovery/calls, bounded argument validation, missing-database behavior, and metadata redaction. The packaged collector smoke test also ingests a synthetic event and queries its domain over stdio.
Launch through the explicit /var/jb/bin/sh interpreter; direct exec of a new shell script was rejected by iOS on the development device. Tool discovery through the explicit interpreter was verified.
Saved reviews appear in the app’s AI Insights tab. Publish validates evidence exists and matches the window/app; it rejects duplicate IDs with changed content. It never modifies recorded events or monitoring settings. Reviews and their limited metadata evidence snapshots are retained for seven days after creation. Author is a client-provided label, not authenticated identity.