SwarmMarshal
Your messages sync in
Email over IMAP and OAuth, plus iMessage, Slack, Telegram, and Discord — into a local database on your computer. Nothing is uploaded anywhere.
SwarmMarshal · free desktop app for Windows & Mac
Give it your email. Get answers.
SwarmMarshal is a full mail and chat client — Gmail, Outlook, iMessage, Slack, Telegram, Discord — with an assistant that answers questions from your real messages. “What did the contractor quote me last fall?” gets the answer and the original email. All of it on your machine.
Windows & Apple Silicon Mac
Local-first
Source-cited answers
You pick the model
Keep your current mail client
What it does
Everything your messages know, put to work.
Email and chat go in. Answers, timelines, briefings, and a clean inbox come out.
Ask your history
A source-cited answer from your real messages — not a guess.
Search by meaning
Find the message even when you don't remember the words.
Every inbox, one client
Gmail, Outlook/IMAP, Apple Messages, Slack, Telegram, Discord — works alongside the mail client you already use.
Build a timeline
Dated events and decisions from real messages, with sources.
Commitments don't slip
Promises buried in email surface as calendar items and tasks.
Hire AI workers
Real hires in their own supervised process, on your machines — and every paired computer shares the load.
How it works
The magic ingredient is context.
A chatbot starts every conversation knowing nothing about you. SwarmMarshal does the opposite: it turns your own messages into durable knowledge, keeps it on your machine, and hands the right slice to the AI at the right moment.
Knowledge gets extracted
As messages arrive, a background pipeline pulls out the durable stuff: people, companies, commitments, deadlines, decisions. Every fact keeps a link to the message it came from — nothing is asserted without a source.
A question assembles a context pack
When you ask, hybrid keyword + semantic search selects the relevant messages and facts — not your whole life, just the slice that question needs.
The model answers from the pack
Whichever AI you chose — local or cloud — answers only from that grounded context and cites its sources. If the record doesn't contain the answer, it says so instead of guessing.
Why this beats a chatbot
You never re-explain who Acme is or dig up the thread yourself. The context compounds on your machine the longer you use it — and every answer can be audited by clicking through to the original message.
Models are swappable. Context isn't.
Because the knowledge lives with you, not the vendor, you can point it at a new model the day it ships — local or cloud — and it answers like it has known you for years.
Real screens · demo profile
This is the actual app.
Captured from a real profile, not a concept deck.
The knowledge it builds
People, promises, decisions, and relationships extracted from your own history — each fact linked back to the message it came from.
Search by meaning
Semantic and keyword search across every channel — find the message even when you don't remember the exact words.
A real, fast inbox
Email, iMessage, Slack, Telegram, and Discord in one clean client. Everything it touches becomes searchable history.
Your day, briefed
The Today dashboard: what changed, who's waiting, deadlines, and what needs attention — with sources behind every fact.
Calendar & commitments
Google and Microsoft calendars, plus deadlines and promises pulled from email — still linked to the original message.
You pick the models
Local models via Ollama or LM Studio next to cloud keys. Per-task routing, budgets, and a fully local option.
Your data, your rules
Private by default. Powerful when you ask.
Your mail — and the knowledge built from it — stays on your machine. Each task goes to the right AI: a free local model for the routine and the private, a bigger model only when you say so.
Run fully local if you want
Every AI job in the app — sorting, extraction, search, even the assistant — can run on a free local model through Ollama or LM Studio. No cloud key required, ever.
A privacy floor the router enforces
Every AI call carries a privacy class. Work pinned local-only can never fall back to a cloud model — if no local model is available, it stops and tells you rather than quietly sending data out. Subscription CLI routes like Claude Code don't count as local, and the router knows it.
Costs that can't run away
Monthly budgets, a hard spend guard, and a message pipeline that prefers $0 routes — paying per message for routine mail processing is off unless you explicitly turn it on.
For the technical reader
The parts reviewers usually ask about.
It gets more interesting under the hood.
Models are benchmarked on your hardware
Model Scout runs calibration suites — built from the app's real production prompts — against the models your machine can actually hold. Routing preferences are gated on those scores, fail-closed: an unproven model doesn't silently take over your mail pipeline.
Local benchmarks →It's an MCP server too
Point Claude Code, Codex, or any MCP-capable agent at your communication history. Your agent gets the same source-grounded search and context packs the built-in assistant uses — behind an explicit approval gate.
The MCP surface →The architecture is documented
Message pipeline, knowledge graph, context engine, routing, and the trust model — written up properly for people who want to verify the claims rather than take our word.
Technical white paper → Architecture overview →
Try it properly
Kick the tires in ten minutes.
The fastest honest test of whether this is real:
Install and connect one inbox
Gmail, Microsoft, or IMAP — the OAuth walkthrough takes a few minutes, and the first sync starts immediately.
Ask something only your mail knows
“What did the plumber quote me?” “When does my lease renew?” Then do the thing no chatbot lets you do: click the citation and read the original message behind the answer.
Switch the model out from under it
Flip the assistant to a local model via Ollama and ask again. Same context, same citations, zero cloud. That swap is the whole thesis in one click.
Part of the Swarm family
Better with the rest of Swarm.
SwarmMarshal CRM →
Ask across your work inbox and CRM in a single question, and see the team's daily agenda on your Today dashboard. Personal accounts stay separate.
Vibes →
Describe little tracker apps in plain English — the built-in Vibes surface can reference your contacts, messages, and calendar.
SwarmSpan →
Your assistant finds SwarmSpan on its own — moving a file or the clipboard to your other machine becomes part of the conversation.
Free to try
Install it, connect your accounts, ask your first question.
Every answer links back to the real message. Your data stays on your machine.