Reads Apple Calendar and Apple Contacts read-only to pull calendar events and merge contact identities into one person record across email, phone, and name.
Mirrors Gmail to a local notmuch-indexed Maildir, enabling local search, relationship context, and saving Gmail drafts (including reply drafts) without sending or modifying mail.
Takes a read-only snapshot of the local iMessage database to include text-message history, per-channel recency, and exact-phrase search in the relationship mesh; requires Full Disk Access.
Uses locally running Ollama with nomic-embed-text to provide semantic search over the relationship mesh, keeping embeddings and queries on the local machine.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@exo-mail mcp What did I promise Jordan in our last thread, and what's still open?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Walk into every conversation with the whole relationship in front of you, not the fragment one inbox remembers.
Four systems each hold a quarter of every relationship you have, and nothing joins them. So the questions you actually ask — when did I last talk to this person, what are our open items, where did we leave off — cost four searches and a guess.
exo-mesh answers them in one local read. It mirrors your Gmail into a notmuch-indexed Maildir, pulls your calendars, takes a read-only snapshot of iMessage, merges your contacts, and resolves the whole pile into a mesh: one row per person, carrying every address and number they own, per-channel recency, who usually starts the conversation, and how many years it's run.
Everything stays on your disk. No service, no daemon, no API key on the query path. Your relationship history is the most revealing corpus you own, and this is the copy that answers in single-digit milliseconds without renting it to anyone.
And it's the store an agent can read. exo-mail mcp speaks MCP, so every agent, skill and draft
you build reasons over the whole relationship instead of one inbox's fragment, from a file on your
disk under a retention policy you wrote. It runs against a six-figure message archive, not a demo
fixture.
Four channels. One person. Zero egress.
I'm leaving out big parts of my implementation because I'm not interested in disclosing all the specifics of my security and governance setup. You can see parts of what I'm doing here: agentrust-io.com if you're interested. And generally assume this is a project I'm sharing, not a product. I hope you get value. Enjoy!
Related MCP server: Persome MCP Server
Total recall before a conversation. exo-mail contact show "Jordan Rivera"
assembles every channel at once: the last thread, the last text, meetings you have
shared, who tends to start them, how the relationship has trended.
Messages that match reality. Draft a reply and it comes with a context block — what you last discussed, and the things you told them you would do, extracted from your own sent mail. You stop opening with "it's been ages!" to someone you texted yesterday.
Reach out where the relationship lives. The mesh knows your warmest channel with
each person. Formal email can be stone cold while you text every week.
exo-mail reconnect "Priya Shah" points at the channel that is actually alive.
One person, not four strangers. A work address, a personal Gmail and a mobile number resolve to a single identity, so "who is this?" has one complete answer.
Get told who you're losing before they're gone. exo-mail contact drift --days 120 reports who has gone quiet across every channel — not the
email-shaped shadow of it. No false alarm about someone you saw on Tuesday.
Know which relationships outlived the job that made them. BEDROCK: seven-plus years of span, two-way in at least one channel, still alive in the last eighteen months. The tool computes those three; the fourth — did it survive a context change? — is a human call it records rather than guesses. See docs/01-CONCEPTS.md.
Find it whether you remember the words or only the gist. exo-mail search
for a notmuch query, exo-mail semantic for meaning (local embeddings via
Ollama — nothing is sent anywhere), and exo-imsg-search
for an exact phrase you remember from a text three years ago.
Open one thing in the morning instead of five. exo-mesh composes today's
meetings, your task counts, the mail that needs a reply, and whether you've written a
journal note yet — one read across the local stores.
No tool in this repository sends mail, marks spam, or moves anything to trash. Not the CLI, not the MCP server. There is no flag for it. The only outbound verb is save a Gmail draft — including the reply lane — and a human presses Send in Gmail.
Everything the tools write is reversible: notmuch tags in the exo/* namespace,
the matching Gmail labels, local SQLite databases you can delete and rebuild, and
drafts. Sources are read-only: your iMessage database, Apple Calendar and Apple
Contacts are opened read-only and never written.
Read docs/03-SAFETY.md before you run anything that says
--apply. Two tools can create or modify records outside their own store
(exo-mail-ensure-labels creates Gmail labels; exo-mesh-bedrock-apply writes
markdown person files), and both are documented there with their gates.
macOS, Python 3.11+, and roughly: brew install notmuch, a lieer checkout per
account, pip install -r requirements.txt into a venv, one notmuch setup,
ollama pull nomic-embed-text if you want semantic search.
The full, tested, line-by-line version — including the Google OAuth steps and the scheduled sync — is docs/00-INSTALL.md. Do that one instead of this paragraph.
git clone https://github.com/AaronRoeF/exo-mesh.git
cd exo-mesh
python3 -m venv venv && ./venv/bin/pip install -r requirements.txt
./bin/exo-mail-python bin/exomail_config.py # print what every setting resolved toYou almost certainly need no configuration file: your addresses, your name and your
mail root are read from notmuch config, which you set up anyway.
config.env.example documents every knob for when you do.
Prerequisites, venv, OAuth, first sync, scheduling | |
The four channels, identity resolution, BEDROCK | |
Every command, grouped by what you are trying to do | |
What writes, what cannot, and every gate | |
Every setting, how it resolves, worked example | |
Where this does not work, honestly | |
Every command and what it gets you — the reference. | |
Seven things a local relationship mesh makes possible — five that work, two that do not exist yet. | |
Reading this as an agent? Start there — MCP tools, JSON contract, exit codes, what writes. |
macOS only. Gmail only. Single writer — run the sync on exactly one machine. Reading iMessage requires granting Full Disk Access, and without it the database reads zero rows with no error, which is why every path asserts a non-empty result rather than succeeding quietly. Semantic search needs Ollama running locally. A large mailbox costs real disk. Details and the rest of the list: docs/05-LIMITATIONS.md.
exo-mesh is one subsystem of a larger local-first setup. Each of these stands alone; you do not need any of the others to use this.
The layer this plugs into — memory, skills, hooks and MCP servers that compound across sessions. exo-mesh is where its relationship context comes from. Start there if you want the whole environment rather than the mail and relationship half. | |
The field-tested patterns this was built with — what worked, what did not, and why. Read it if you want to build something like this rather than run this. | |
The same local-first shape applied to books instead of people: your highlights become one short lesson a day, on your own disk. |
Working software that one person runs every day, published because the pattern is worth copying — not a product, and not something anyone is on call for. Expect to read the source. Issues and forks welcome; support is not promised.
MIT. See LICENSE.
This server cannot be deployed
CRM, relationship intelligence, and multi-agent orchestration for AI agents.
Personal memory layer: files, notes, messages, tasks and email, semantically indexed.
Query one person's second brain: contacts, companies, notes, files and connected accounts.
An AI-first personal CRM you run in natural language: contacts, reminders, notes, and more.
