MailDrop

A desktop app that bridges email and AI. No OAuth. No IT approval. Works anywhere Outlook does.

← All case studies

The problem

Every AI assistant available today can read files, write code, search the web, and query databases. None of them can read your Outlook inbox.

The reason is architectural. AI tools connect to external systems through a protocol called MCP — Model Context Protocol. There are MCP servers for Snowflake, Jira, Confluence, Slack, GitHub, and dozens of other systems. There is no MCP server for Microsoft Outlook or Office 365. Microsoft hasn't shipped one. As of April 2026, there's no public timeline for when they will.

This means the single application most knowledge workers live in all day — email — is a black box to AI. Your AI can write a SQL query against a petabyte data warehouse, but it can't read the email asking you to do it.

Meanwhile, on this site we've been building communication patterns that replace meetings with structured, written, AI-assisted processes — Tunnels, Pulses, Forges, and Senates. These patterns all produce markdown files. They all depend on an AI being able to read incoming communication and help draft outgoing responses. They all assume a bridge between email and AI.

The bridge doesn't exist. So we built one.

What MailDrop is

MailDrop is a desktop application that converts email files into clean markdown. It has no AI inside. No API keys. No cloud calls. No telemetry. It's a Python app with a dark UI that sits in a narrow strip next to Outlook.

You drag an email from Outlook to a Finder/Explorer folder. MailDrop detects the new file, strips the noise — signatures, legal disclaimers, "Sent from my iPhone", HTML formatting soup — and produces a clean .md file with the subject, headers, and body in structured markdown.

That's it. The .md file is the bridge. You hand it to whatever AI you have — Claude, ChatGPT, Copilot, Gemini, your corporate portal — and now the AI can read your email and help you respond.

How it works

The inbox and outbox

MailDrop watches a folder called inbox. When an .eml file appears, it parses it and writes a clean .md file to inbox/outbox. The original stays in the inbox. The conversion goes to the outbox. One Finder window, up and down.

• Drag from Outlook to Finder — this creates a .eml file. Drop it in the inbox folder (pin it to your Dock for one-drag access).
• MailDrop converts automatically — the drop zone flashes green, the preview pane shows the clean markdown, the outbox has the file.
• Copy Path or Copy Markdown — hand it to your AI.
• OK — reset, ready for the next one.

Note: direct drag from Outlook to MailDrop won't work on most systems. Outlook doesn't hand off .eml data to non-native drag targets. The inbox folder is the workaround. It's one extra step, but it's reliable on both macOS and Windows.

Noise stripping

Corporate email is 40% content, 60% noise. MailDrop strips:

• Email signatures (-- delimiter and everything below)
• Legal disclaimers ("This email is intended only for the named recipient...")
• Device footers ("Sent from my iPhone", "Get Outlook for iOS")
• HTML formatting → clean markdown with headers, bold, links, tables preserved
• Excess blank lines and trailing whitespace

What survives: subject line as a heading, From/To/CC/Date headers, and the actual message body. The stuff you need. Nothing else.

The four meeting patterns

MailDrop isn't just an email converter. It also generates starter files for the four communication patterns described in the meeting reform case study:

Each pattern generates a .md file with an AI instruction block at the top. Any AI that reads the file knows exactly what the pattern is, what rules to follow, and how to help. The AI doesn't need to know your system — the file teaches it.

Thread IDs

Every pattern file is named with a thread ID:

TID_20260413_847291_FORGE_migration-plan.md

Five fields, split on underscore:

• TID — prefix marking this as a tracked thread
• 20260413 — date
• 847291 — random six digits for uniqueness
• FORGE — which pattern
• migration-plan — human-readable topic

The same TID appears inside the file. All files in a conversation share one TID. Any AI on any end reads all files with that TID and reconstructs the full history. No memory needed. The files are the memory.

The AI handoff

Using it with Claude Code

The smoothest path. Claude Code can read and write files directly:

You:   "Read ~/Documents/maildrop/inbox/outbox/TID_20260413_847291_TUNNEL_q2-status.md
        and draft a response. Same format, same thread ID."

Claude: [reads the file, writes a response to disk]

No clipboard. No copy-paste. The AI reads the clean markdown, understands the thread ID, and produces a response you attach to an email.

For a Forge synthesis:

You:   "Read all files in my outbox that contain FORGE in the filename.
        Produce a Forge Report: cluster ideas, surface outliers,
        map conflicts, recommend a shortlist."

Claude: [reads all submissions, produces synthesis]

Using it with any other AI

You don't need Claude. You don't need Claude Code. Any AI that can read text works.

1. MailDrop creates the .md file in your outbox
2. Hit Copy Markdown in MailDrop
3. Paste into ChatGPT, Copilot, Gemini, or whatever your IT department allows
4. Type your instruction: "Draft a response to this email. Professional tone. Address the three questions."
5. Copy the AI's response back into Outlook

The clipboard is the bridge. It's one extra step compared to Claude Code's direct file access, but the result is the same: the AI reads clean structured content instead of HTML soup with six nested reply chains and a legal disclaimer.

Configuration

MailDrop reads a maildrop.toml file next to the app:

[paths]
inbox = "~/Documents/maildrop/inbox"
outbox = "~/Documents/maildrop/inbox/outbox"

[window]
width = 340
always_on_top = true

[user]
name = "Your Name"

Edit the paths to match your system. Set your name so it pre-fills in the UI. That's the entire configuration.

Get it

MailDrop is free. No license. No registration. No usage limits. Take it, use it, share it.

What's inside

One Python file. ~400 lines. PySide6 for the GUI, html2text for HTML-to-markdown conversion, Python's standard library email module for .eml parsing. No external AI libraries. No network calls. Everything runs locally.

80 automated tests verify the parsing, noise stripping, thread ID generation, and all four meeting patterns.

Why this exists

The ideal solution is an MCP server for Microsoft 365 — Claude reads your inbox directly, drafts responses, sends them. No drag-and-drop. No clipboard. No intermediary app.

That server now exists. The community built it: ms-365-mcp-server on GitHub. If your organization's Microsoft 365 tenant permits third-party OAuth app registrations, it's the better path. One install command and Claude has direct inbox access.

Most enterprise tenants require IT or admin approval before a third-party app can connect to Microsoft 365. That process takes time, and in some environments it doesn't happen at all. MailDrop is for those environments. No OAuth. No app registration. No network calls. Drag an email to a folder, get clean markdown. It works on day one regardless of what your IT policy says.

MailDrop is deliberately simple — a file converter and template generator with no AI inside — because the AI is wherever the user already has it. Some people have Claude Code. Some have ChatGPT. Some have a corporate Copilot license. MailDrop doesn't care. It produces clean markdown. Any AI can read clean markdown.

The meeting patterns survive regardless of the transport. The Tunnel protocol works whether the files are produced by MailDrop, by an MCP server, or by someone typing into Notepad. The tool is disposable. The patterns are not.

The human/AI split

• The human identified the gap. No MCP server for Outlook. Email is a black box to AI. The meeting patterns need a bridge. That's observation, not computation.
• The human designed the workflow. Inbox/outbox structure, TID naming convention, pattern templates with embedded AI instructions, the decision to keep AI out of the app itself. Those are architecture decisions.
• The AI built it. The Python app, the GUI, the noise stripping logic, the packaging, the tests, this page. One session. Maybe two hours total, including testing and iteration.
• The human tested it live with real Outlook emails and caught every UX problem — overlapping UI elements, missing feedback during conversion, the Outlook drag-and-drop limitation, filenames that don't split cleanly. The AI fixed each one in minutes.

The pattern: human sees the gap, designs the solution, tests the result. AI writes the code, the docs, and the deployment. Neither could do the other's job.

← Back to case studies