Your Mac runs the engine
The autonomous agent daemon lives here. Work loops, worktrees, tool calls, and your Claude key all execute locally, within your machine's limits. Heavy compute and private data never leave the building.
SELF-HOSTED AI FACTORY & BRAIN
Factory is an autonomous agent engine you host yourself. It ships PRs, runs work loops, and asks before it does anything sensitive — while the heavy compute and every credential stay on your machine. Watch it, steer it, and command it from native iPhone, Mac, and web apps.
Bring your own Claude· No metered middleman· Open, inspectable code
to: client@example.com · subject: Your July receipt
Drafting winback email variant B
Wiring events feed to store
THREE TIERS
The design keeps the expensive, private parts on your machine and makes only a thin, replaceable relay touch the cloud. Nothing about your work is trapped in someone else's account.
The autonomous agent daemon lives here. Work loops, worktrees, tool calls, and your Claude key all execute locally, within your machine's limits. Heavy compute and private data never leave the building.
A small stateless worker passes read-model state one way and commands the other. It holds no compute and no secrets — it's a mailbox, tenant-scoped and fail-closed, that you can swap or self-host.
Native iPhone and Mac apps plus a web client render the live factory, surface what needs you, and let you command the engine — a chat message, an approval, a nudge — from wherever you are.
Commands are the only write path — a single, auditable INSERT the engine picks up.
WHAT YOU GET
Daemon health at a glance, a Needs You queue for approvals, the live task list across every department, and the PR pipeline — all in one scroll. Tap a task to see its branch, budget, runtime, and current action.
Sending an email, merging a PR, running a shell command — sensitive tools pause and wait for your Approve or Deny. Nothing irreversible happens behind your back.
The read-model refreshes on a tight poll, so tasks advancing, PRs landing, and events firing show up on your phone seconds after they happen on the Mac.
The engine calls Claude with your key. No token markup, no metered middleman between you and the model — just your account, your rate limits, your bill.
Log in to your Brain to pair a device; the broker mints a short-lived, tenant-scoped token per request. Row-level security is fail-closed — an unpaired client simply sees nothing.
Compute, credentials, and source live on your Mac. The cloud only ever sees a tenant-scoped read-model and the commands you send — never your keys, never your repos. If the relay vanished tomorrow, your factory keeps running.
THE BRAIN
The factory ships work; the Brain remembers why. Memories, files, and live tmux sessions the engine uses all get a proper UI — so the context your agents run on is something you can read, search, and steer, not a black box.
$ brain ls memories/ 142 notes · updated 31s ago › project-gtm-loop-runbook.md › reference-prod-deploy-webhook.md › home-purchase-shortlist.md files/ receipts, decks, exports › receipt-jul.pdf › icna-deck-v4.key sessions/ 2 live tmux gtm-loop attached · 14m product-brief attached · 4m $ ▍
WHY SELF-HOST
Built for people who'd rather own the machine than rent the dashboard:
Clone the repo, point the engine at your Claude key, and pair a device. It runs on Mock data out of the box — flip one switch to go live against your own backend.
Mac build coming soon · the web & iOS clients are in the repo today