Email Checker: Local AI Email Assistant for Mac

TL;DR

A developer built Email Checker, a local macOS app that scores incoming emails by priority and drafts replies using your local LLM—eliminating inbox distraction without cloud dependencies or SaaS fees.

The Problem: Email as Context Killer

Lorentz Yeung identified a core productivity drain: not email volume, but switching costs. Notifications pull deep work into surface-level replies, fragmenting focus. Existing solutions demand either SaaS subscriptions or trusting email to third-party servers. He wanted something local, automated, and unopinionated—a small tool that solved one problem exceptionally well.

How Email Checker Works

The app runs on a cron schedule (default: every 15 minutes) and performs a four-step pipeline:

• Fetch: Queries Mail.app inbox via AppleScript for unread messages
• Score: Ranks emails HIGH/MEDIUM/LOW using keyword triggers (urgent, approve, feedback) and sender trust weights
• Enrich: For HIGH-priority emails only, fetches up to 10 prior messages in the thread to provide context
• Generate: Sends thread history to a local LLM to draft contextual replies; generates formatted email report with previews and drafts

The critical design choice: drafts are generated only for high-priority emails, reducing hallucination risk and avoiding generic responses. Low-priority emails (GitHub notifications, newsletters) receive previews only. Everything is marked read after reporting, clearing notification noise.

Why This Matters for Developers

Email Checker demonstrates a shift in how developers think about AI tooling. Rather than delegating to cloud APIs, Yeung built an on-device, privacy-first assistant that composes naturally with existing infrastructure (macOS Mail.app, local LLMs like Ollama or LM Studio, OpenClaw's scheduling system).

The app also validates ClawHub, OpenClaw's emerging skill marketplace. Publishing a single-purpose tool to a shared registry lowers the bar for developers to share composable local agents, moving beyond monolithic platforms toward modular, user-owned AI infrastructure.

Technical Architecture

The setup prioritizes accessibility over complexity. A bash wizard auto-detects Mail.app accounts, prompts for LLM provider selection (LM Studio, Ollama, OpenAI, or none), and writes configuration in ~2 minutes. Users grant Terminal automation permissions via System Settings—a single macOS security step—and cron handles scheduling thereafter.

LLM flexibility is baked in: developers can use fully-local models (Ollama), remote self-hosted setups (LM Studio), cloud APIs (OpenAI's GPT-4o), or disable draft generation entirely for keyword-only prioritization. This modularity avoids vendor lock-in and respects compute constraints.

Reports arrive as formatted emails (not in-app notifications), preserving the principle of not interrupting your workflow. Users can approve and send drafts via a Telegram integration with their OpenClaw agent, completing the loop without touching email clients.

Limitations and Design Trade-offs

Email Checker is macOS-only, tying it to Mail.app and AppleScript—a deliberate choice for depth over breadth. No Linux or Windows support. No native mobile interface; reports arrive as emails, requiring manual Telegram-based approval for sends. The keyword-scoring system is static and user-configured; it lacks dynamic learning from reply patterns (though Yeung lists this as a potential future direction).

The LLM drafting is context-limited to thread history only, not calendar or metadata signals, though Yeung explicitly flagged this as a planned enhancement. For teams requiring audit trails or compliance, the lack of cloud logging is a feature; for enterprises, it may be a gap.

Why OpenClaw as the Foundation?

Yeung had already been running OpenClaw as an always-on local assistant for Telegram integration, task scheduling, and workspace management. Email Checker emerged as a natural skill—a focused, composable module that leverages OpenClaw's scheduling primitives and agent ecosystem. Publishing to ClawHub (the official skill registry) democratizes the tool: anyone running OpenClaw can install it in one command without forking repos or managing dependencies manually.

This reflects a broader philosophy in the open-source AI ecosystem: local-first, composable tools beat all-in-one platforms. openclaw-board-installer Building-OpenClaw-Skills

Roadmap and Community Signal

Near-term improvements include explicit approval workflows (queuing drafts before sending), daily digest mode (one summary instead of periodic reports), and richer priority signals (calendar context, CC patterns, sender history analysis). The author explicitly invites community ideas and showcase of other OpenClaw projects, signaling openness to contributions and integration.

Practical Setup Path

Installation is three steps: clone the repo, run the wizard (auto-discovers Mail accounts and LLM config), grant Terminal automation permissions, and done. The first cron fire produces your first report. Documentation includes setup walkthrough, workflow diagrams, and config reference—removing friction for adoption.

Key Takeaways

• Email Checker eliminates inbox context-switching by scoring emails offline, drafting replies locally, and delivering a single prioritized report—proving that AI assistants solve friction, not just capability gaps
• Local-first, single-purpose tools published to composable registries (ClawHub) represent a sustainable alternative to monolithic SaaS and vendor-dependent platforms
• macOS Mail.app + AppleScript + local LLM integration demonstrates that privacy-respecting AI workflows are now achievable without cloud dependencies or subscription overhead
• The project validates OpenClaw's skill-marketplace model: developers can publish focused, user-installable agents that compose naturally with existing local infrastructure
• Yeung's roadmap (approval flows, digest mode, richer signals) indicates iterative, community-driven development—suggesting Email Checker is a prototype for a new class of local AI tools

Source: Lorentz Yeung, Medium, March 2026. Published to ClawHub. macOS only.