Cortland 🍎

Your agent brings the brain. Cortland brings the hands — and the human gate.

Give an AI real access to your Mac — Mail, Calendar, Reminders, Notes, files — without giving it the keys. Every consequential action previews first, waits for your approval through a channel the model can't touch, gets logged, and can be undone.

Cortland has no model of its own, no cloud, and no web access. That is the point. Whatever agent you already use — Claude, Codex, Cursor, or a local Gemma that never leaves the machine — gets a governed way to touch your real data and to reach you.

A worked example

An agent watched a handful of retailer pages for days, waiting for a limited-run controller to come back in stock. When it finally did — unattended, at an hour I was nowhere near my desk — my phone buzzed with an iMessage. I bought one.

Cortland never touched a website. It has no web-fetch tool and never has. The watching was done by a coding agent with internet access. Cortland was the part that reached me: on my own Mac, through Messages, with the recipient fixed in local config so nothing the agent generated could point that message at anyone else.

That division is the pitch in one story. Your agent has the reasoning and the internet. Cortland has your Mac, your Apple apps, and your attention — and every outward action still stops to ask.

The part worth telling

The first version cheated, and I only found out by reading my own audit log.

No governed "tell the human" tool existed. The iMessage bridge could answer me, but nothing could start a conversation — so the monitor imported Cortland's send module directly and called it. It worked, and it was genuinely Cortland doing the sending: same owner handle from config, same AppleScript escaping, same hourly rate cap. But it went around the framework instead of through it, so days of running left zero audit rows.

That is this project failing its own test. A guarantee with no legitimate path through it is a guarantee people route around — and the person who routed around it was the one who wrote the guarantee.

The fix is notify_owner: write-safe, because a message to you is not an outward action, it is the review loop. It has no recipient argument at all — the handle comes from local config, so no model output and no injected content can redirect it. Its hourly cap is counted from the audit log rather than from memory, so a cron job that starts fresh every run still can't exceed it. A notification now leaves a row behind. This one is real, copied out of the audit DB after sending a test message to my own phone:

notify_owner | Messages | write-safe | undo=none | dry_run=0 | ok
principal: cli:notify   tool_version: 0.2.0
args: {"text":{"redacted":true,"length":84,"sha256":"3e389e33…"},
       "source":"cortland-test"}

The body is stored as a length and a hash — the row proves a message was sent without keeping what it said. source stays readable because it is a label, constrained to a slug, and not content anyone can compose.

Related MCP server: orchard-mcp

What it looks like in use

Text it like a contact — the bridge, answered by a 4 GB model on an 8 GB MacBook Air. Nothing left the Mac:

You:       Add a reminder to call the vet tomorrow at 2pm
Cortland: Received — working on it…
Cortland: Added "Call vet" for tomorrow, Aug 3 at 2:00 PM.

The texting interface needs a second Apple ID, signed into Messages on the Mac only — a mouthpiece, not a worker. It owns no mail, no calendar, no data. You text it from your phone like any contact; the Mac still does the work as you. Your system iCloud stays yours. Claude Code, Cursor, Codex, and other MCP clients skip this step: they talk to the same tools over stdio.

Ask about your week — reads are free, no approval needed:

You:       Anything from the school this week?
Cortland: Three emails from Lincoln Elementary. Two are the weekly
            newsletter; one from Ms. Alvarez on Tuesday asks for a
            permission slip by Friday.

Ask it to change something — writes stop and ask, every time:

The code in that prompt is generated by the framework and never enters the model's context — so even a fully prompt-injected model can't forge your approval. Ignore the message and the action expires, refused.

Ask what it can do — answered from the tools actually installed, so it can't overclaim:

You:       what can you do?
Cortland: Here's what I can do with your Mac:
            • Calendar: list, look up, create, delete
            • Mail: search, read, follow threads, draft, send (asks first), flag, file
            • Notes: search, read, create, add to
            • Reminders: list, search, create, tick off, delete

            Anything that deletes or changes things asks you first — I text
            you a code and wait for "yes <code>". Reads just happen.

In Claude Code or Claude Desktop, the same tools are just there:

You: send a reply to Ms. Alvarez saying the slip is coming Thursday Claude: [searches mail, reads the thread] I can send this from your Personal account to alex@example.com. Approve in the usual channel.

mail_send is write-gated. Live mode plus per-action human approval. A stolen token can only ask.

Requirements

macOS only. Nothing here runs on Windows or Linux — the tools talk to Apple apps via Automation.

iMessage bridge only: macOS 13+ (modern chat.db), Full Disk Access for the process that runs Node, Automation → Messages, and a second Apple ID signed into Messages.app — not into System Settings. See SETUP.md §5.1.

Osaurus as a client: Apple Silicon, macOS 15.5+.

Not required: a developer account, iCloud for Path A, or any API key.

Quick start

Fastest path to “it searched my mail” — Cursor (or Claude Code) as the brain, no second Apple ID, no Ollama:

xcode-select --install          # if `xcode-select -p` fails
git clone https://github.com/cortland-suite/cortland.git
cd cortland
npm install && npm run build    # first install compiles sqlite — a minute

Then Cursor Settings → MCP, add servers pointing node at packages/mail/dist/server.js (and reminders/notes/calendar if you want them). Absolute paths. JSON is in docs/08_local_models.md.

Ask: search my mail for anything from school this week. macOS will prompt Automation (Cursor → Mail). Allow it.

Writes stay previews until you set "live": true in ~/Library/Application Support/cortland/config.json. Even then, send/delete asks you per action.

Texting it from your phone is a longer path (second Apple ID + Ollama). SETUP.md is the ordered list; docs/08 is the illustrated one.

Why it exists

Most MCP servers for personal data are thin wrappers around AppleScript: delete_email, send_message, executed the instant a model calls them. That's a hard thing to trust with an inbox, and the alternatives don't help — cloud assistants (Poke, Arlo, Lindy) can't touch Apple-native data at all, because Apple gives them no API, and self-hosted agents like OpenClaw run ungoverned (a CVSS 8.8 in January 2026, with permission gates still on the roadmap).

Cortland inverts the default. A tool call previews what it would do unless you've opted into live mode, and even then every consequential action waits for you. The framework enforcing that is a small library with a test suite proving each guarantee can't be bypassed — and every tool in the suite is built on it.

The guarantees

Install

From the repo:

git clone https://github.com/cortland-suite/cortland.git
cd cortland
npm install
npm run build
npx cortland setup

The wizard asks before every step and records what it did. Full walkthrough — including the iMessage bridge, permissions, and model choice — in SETUP.md. Connecting Cursor, Codex, LM Studio, Osaurus, or Ollama: docs/08_local_models.md.

Packages

Bring your own model

Cortland is the tools. You pick the brain.

Path A. Cursor, Claude Code, Claude Desktop, Codex, LM Studio, or Osaurus hosts the model. Cortland is an MCP plugin. No second Apple ID.

Path B. You text it. A second Apple ID signs into Messages on the Mac (mouthpiece only). Ollama runs Gemma 4 on disk. Approvals are yes <code> in the same thread.

Field-tested on an 8 GB M2: gemma4:e2b-it-qat (4.3 GB) makes clean tool calls and refuses honestly. The governed contract matters more with a small model, not less.

Walkthroughs, including LM Studio and Osaurus from a clean install: docs/08_local_models.md. Clean-Mac order, permissions, and the iMessage second-ID steps: SETUP.md.

Privacy model

• Pointers, not copies. The context layer stores metadata referencing messages by ID; bodies are never stored. Delete a message in Mail and the pointer dangles and gets pruned.

• Model use is opt-in and declared. Works fully deterministically with no model. Configure one and its network egress is declared in config and recorded in the audit log on every run.

• Nothing phones home. The only optional outbound call is a push ping you configure yourself, and its body is a fixed string carrying no information.

Docs

• SETUP.md — clean Mac to working assistant, with the gotchas.

• docs/08_local_models.md — connect Cursor, Codex, LM Studio, Osaurus, or Ollama; illustrated, from zero.

• SECURITY.md — reporting, and what's in scope.

• docs/ — one design doc per component: the framework contract (01), Mail (02), folder-as-API (03), the context layer (04), remote access (05), the iMessage bridge (06), the threat model / review guide (07), and local models (08).

• NOTES.md — the engineering log: decisions with dates, open questions, and every field finding, including the ones that were embarrassing.

License

MIT. See LICENSE.