The Desktop

One terminal. Three panes. Two Claudes. No special connectors required.

← All posts

The problem everyone has

Your AI can read files, write code, search the web, and query databases. It cannot read your email.

There's a protocol called MCP (Model Context Protocol) that lets AI tools connect directly to other software — databases, calendars, file systems. Think of it as a universal adapter between AI and everything else. Most email clients don't have one. Your AI is brilliant and your inbox is a black box.

But Apple Mail exposes itself to AppleScript — a simple automation language built into every Mac. Your AI can write AppleScript. Which means your AI can read, search, compose, and send email without any special connector at all.

You could wait for Microsoft or Google to build an official AI connector. Or you could use what already works.

The layout

The left half is a terminal — a text-based window where you type commands and talk to AI. The app is called iTerm2, a popular terminal for Mac.

It's set to "float," meaning it stays visible on every macOS virtual desktop. macOS lets you create multiple desktops — separate workspaces you swipe between with a three-finger swipe on your trackpad, or by pressing Control + Left/Right arrow. To set this up on any Mac app: right-click the app's icon in the Dock, go to Options, and choose All Desktops. For iTerm2 specifically, you can also set it in Profiles → Window → Space → All Spaces.

Multiple desktops, one AI

The terminal follows you everywhere. The right half changes depending on the job. I use three desktops:

• Communication — Finder, Calendar, Apple Mail, and Teams. This is the layout shown above. Email comes in, Claude reads it directly, responses go out.
• Editing — Chrome with Google Analytics and a site preview, plus a Finder window for navigating files. This is where I write and publish the YOU++ site.
• Day job — Chrome with our Jira board and Snowflake (a cloud database), plus a text editor for scratch notes. This is where I work data operations tickets with my team.

The AI is the same in every desktop. I don't close it and reopen it when I switch tasks. I don't start a new conversation and re-explain who I am and what I'm working on. The context stays loaded. When I swipe from communication to editing, the terminal is already there with everything it knew five minutes ago.

This is why the floating terminal matters. Without it, every desktop switch is a reintroduction meeting — "Hi, I'm Bill, here's what I'm doing, here's what I need." With it, you pick up mid-sentence. The AI remembers. You just change the scenery.

The left half: always the same

Two Claude Code sessions side by side on top. One plain shell on the bottom.

• Claude Code, session 1 — the current project. Right now it's reading email, drafting responses, and writing this post. Tomorrow it might be synthesizing a Forge or deploying a site update. This is the workhorse.
• Claude Code, session 2 — a different context. Day-job tickets, database queries, code review. Two sessions means you never lose your train of thought when you switch tasks. Each one remembers its own conversation.
• Plain shell (zsh) — no AI. Zsh is just a command line — the text interface where you type short commands to manage files, check on things, or launch programs. I open mail drafts in Sublime Text here, run git, and search the archive. The command line is fast because it doesn't think. Sometimes that's what you want.

Two AIs and a command line. That's the whole left half.

The right half: the communication station

Four windows in a 2×2 grid:

• Finder (top-left) — the maildrop directory. Inbox, outbox, archive, and templates are all visible. When Claude drafts a response, the markdown file appears here. You open it in a text editor, make your edits, tell Claude to send it.
• Calendar (bottom-left) — Apple Calendar in day view. Appointments, deadlines, and the daily routine at a glance. Claude can read this too via AppleScript.
• Teams (top-right) — chat and interrupts. Wide enough to show the contact list and a conversation side by side.
• Apple Mail (bottom-right) — all accounts in one app. Exchange (work), iCloud (personal), Gmail (projects). This is the source and the destination. Claude reads from it and sends through it — no copy-paste, no file conversion, no intermediary.

The flow

An email arrives. You see it in Apple Mail. It needs a real response — not a thumbs-up, an actual answer with thought behind it.

1. Tell Claude: "Read my latest unread email and draft a response."
2. Claude reads it directly from Apple Mail via AppleScript. No drag, no file conversion, no clipboard. It strips the signatures and noise and saves a clean markdown draft in the outbox.
3. The file appears in Finder. Open it in Sublime Text (or any editor). Add your personal notes, adjust the tone, write the parts only you can write.
4. Tell Claude: "Send it." Claude reads your edited file, composes the email via AppleScript, and sends it from the right account.
5. The markdown gets archived — file moved to the archive, full-text indexed in a SQLite database. Searchable forever.

Total time: maybe two minutes for an email that would have taken fifteen. And zero context switches — you never left the terminal.

The AppleScript bridge

Here's what's actually happening under the hood. Claude writes small AppleScript commands and runs them in your terminal. You don't need to know AppleScript — Claude does. But if you're curious, here's what the basic operations look like:

Read unread messages:

Read a specific email's body:

Send an email:

Attach a file:

That's it. No API keys. No OAuth tokens. No cloud service. AppleScript talks to Apple Mail the same way you do — through the app's own interface. macOS asks permission once ("Allow iTerm to control Mail?") and then it just works.

The only requirement: macOS with Apple Mail. If you're on Windows or using Outlook exclusively, see the MailDrop approach below — it solves the same problem with a file-based bridge instead.

Why two Claudes

Context is expensive. Not in computing cost — in your head.

You're deep in a database investigation and an email arrives that needs a thoughtful response. You have two choices: blow up your current conversation or ignore the email.

With two sessions, you handle the email in session 1 and go back to session 2 without losing a single thought. Both Claudes remember where they were.

This is the same principle as the Tunnel — don't let the transport destroy the content. Two conversations, two contexts, zero interference.

What this replaces

The old way: Outlook full-screen, switch to ChatGPT in a browser, copy-paste an email into a text box, wait, copy-paste back, switch again, lose your place, repeat. Seven context switches per email.

This layout has zero. Claude reads the email, writes the draft, and sends it — all from the terminal. The only human step is reviewing the draft in a text editor and saying "send it."

No special connector required

AppleScript is built into every Mac. Apple Mail exposes its full interface to it — read messages, search, compose, send, manage folders, save attachments. Claude Code can run AppleScript commands via the terminal. The only setup is a one-time macOS permission: System Settings → Privacy & Security → Automation → allow your terminal to control Mail.

No API keys. No MCP server. No third-party app. No cloud. The bridge is the operating system itself.

The archive

Every email Claude processes and every pattern file you generate gets copied into an archive organized by year and month:

archive/
  2026/
    04/
      TID_20260414_366096_TUNNEL_sylvia-4xl-cost.md
      TID_20260413_103958_FORGE_ali-gift.md
      TID_20260414_095703_TUNNEL_sandbox-resourcing-inbound.md
      ...

Two naming conventions. Processed emails use a timestamp and the slugified subject line:

YYYYMMDD-HHMMSS_slugified-subject.md

Pattern files — Tunnels, Pulses, Forges, Senates — use a thread ID that stays consistent across the entire conversation:

TID_YYYYMMDD_RANDOM_PATTERN_topic-slug.md

The TID goes in the email subject line. Any AI on any end of the thread can search for that ID and reconstruct the full history from the files alone. The random number ensures uniqueness. The pattern name tells you (and any AI reading the file) what kind of communication it is — a Tunnel, a Forge, a Pulse, or a Senate. The slug is human-readable so you can scan a folder and know what's what.

Next to the archive sits a SQLite database with a full-text search index. Every archived file — subject, sender, date, body — gets indexed the moment it's archived. SQLite is a database engine built into every Mac, every iPhone, every Linux server. It doesn't need a server. It doesn't need the internet. It's a single file on your disk.

The search index uses FTS5 — SQLite's full-text search engine. It's the same technology that powers search in Apple Mail, Firefox history, and countless mobile apps. One query, ranked results, instant.

Compare this to Outlook search. Outlook searches through HTML formatting artifacts, nested reply chains, legal disclaimers, and signature blocks. It finds the word you searched for buried inside markup you never wrote. It returns forty threads with the same subject line. You scroll. You guess. You give up and ask someone.

The archive searches clean markdown. No noise. No formatting soup. The text that went in is the text that comes back. One command, one answer.

Six months from now, someone asks "what did we decide about Sandbox resourcing?" You don't open Outlook. You don't scroll through threads. You tell Claude to search the archive. Same result, in under a second.

The archive is just files on disk. The index is just a database on disk. No cloud. No subscription. No server. If you want to move it, copy the folder. If you want to back it up, it's already in your normal backup. If you want an AI to read your entire communication history and synthesize a brief, point it at the archive directory. Every file is clean, structured, and readable.

Build your own

You don't need my exact setup. You need the principle:

• Keep AI visible. Don't hide it behind a browser tab. Give it screen real estate. It's a coworker, not an app you switch to.
• Separate contexts. Two AI sessions beat one. Context-switching inside a conversation destroys both threads.
• Bridge the gaps. If your AI can't reach a system directly, find the automation layer that connects them. AppleScript on Mac. PowerShell on Windows. A file folder as the universal fallback.
• Put interrupts at the edge. Chat belongs at the periphery. Deep work belongs at the center.

What if I don't use Apple Mail?

The AppleScript approach works with Apple Mail on any Mac. If you're on Windows or use Outlook exclusively, the original version of this workflow used MailDrop — a free desktop app that converts Outlook emails into clean markdown files via drag-and-drop. It still works and it's still available.

Not a Claude Code user? Swap the left half.

The left half of this layout is iTerm2 with Claude Code. That's what I use. But the pattern works with any AI that can read files from your computer. Here's how the left half changes depending on your tool:

The right half stays the same no matter what AI you choose. The bridge is AppleScript on Mac, or the file system everywhere else. Any AI that can run a command or read a file can cross it.

← Back to posts