Hello, world

Every system’s first line is the same: print hello world, prove it runs. Ours runs — on a Mac mini, on Apple Silicon, on your desk.

The interesting question was never whether it runs. It’s on whose terms. This is the first post on this blog, so it should say the terms out loud: what Stallari is, what we refuse to build, and — honestly — how far along it actually is.

What we refuse to build

It’s easier to say what a product stands for by naming what it won’t do. Four lines we don’t cross.

Your intelligence stays yours. There’s a hard custody boundary down the middle of the system. Your notes and corpus are plain Markdown, under whatever sync you already use — Stallari reads them, it doesn’t rewrite them into a database it owns. The system’s own state — agent memory, work records, traces — is encrypted on your Mac and exportable in a signed, portable format whenever you ask. Not “exportable if you file a request.” Yours, today, full fidelity. The best lock-in is none.

Autonomy is always accountable. Overnight, on-event, on-demand — the work still isn’t unsupervised. Every step opens a durable receipt, every consequential tool call passes a risk gate, and the whole run leaves a trace you can read in the morning. Drafts wait for your review; decisions wait for you. We don’t sell “no human in the loop,” because a Chief of Staff that can’t tell when to stop isn’t one you’d trust.

Extend everything, depend on nothing. Your agents reach the tools you already use through MCPs, and packs reference those plugins by contract, so capability is something you add, not something we gatekeep — the community marketplace is already live. The same principle governs the models: Stallari integrates with the frontier — Anthropic, xAI, OpenAI, Google — and runs local models (on-device Apple Silicon; Mistral, Llama, Qwen etc.), mixing providers inside a single workflow. The frontier is an option you plug in, not a dependency you’re captive to.

Secure by construction, frugal by default. Security isn’t a setting — it’s the build: memory and traces encrypted on your Mac, cross-device calls cryptographically signed, tool access least-privilege behind risk gates, every peer token individually revocable. And the meter is always visible — every dispatch counts its tokens and its cents, so a night’s work costs cents, not dollars. One more reason local models earn their place beside the frontier.

More than a Personal Assistant

The shape all of that points at is a Chief of Staff, not a chatbot — something that delegates specialised agents to your tools on a schedule, on an event, or on demand. How far that goes is up to you: you compose it from your own tools and your own packs of skills and workflows. Wired up well, the aim is a day that opens with a briefing instead of an excavation — pre-approved workflows already actioned, their work completed, so you begin from a considered position rather than a blank prompt box. A good Chief of Staff also triages what reaches you: it watches the noise so you don’t have to, and raises a Signal only when something genuinely needs you — bothering you when it matters, and leaving you alone when it doesn’t.

Not just a promise: the signal primitives already ship in the pack spec, so packs can declare the events worth surfacing today. The triage layer that acts on them is on the roadmap.

Underneath, the direction is your agents, your providers, your rules. Workflows are portable YAML — a pack you can read, move, and run anywhere, not a configuration trapped in someone’s cloud. That’s not a someday promise: it’s the embedded dispatch engine we’re building right now.

Further out is federation: your own fabric of agents and devices, able to reach the people you choose through scoped, bound agent-to-agent protocols — reach that extends on your terms, gated the same way everything else is.

The honest phase view

Manifestos are cheap; the roadmap is where you check the receipts. The credible thing to do is admit what isn’t done.

• Shipped (Phases 1–8): the Swift harness, vault, and dispatch engine; agent memory with decay, recall, and consolidation, plus hybrid search across your corpus; a notarised macOS app with scheduled dispatch, digests, and privacy controls; multi-device fabrics with leader election; a plugin marketplace with 20+ MCPs and contract-referenced packs; and zero-config secure mesh between your devices.
• Active (Phase 9): the embedded dispatch engine — mix frontier and local providers in one workflow, run portable packs, route the right agent to the right work, with every dispatch metered in tokens and cents.
• Planned (Phase 10): Open Memory — full-fidelity export of every memory and association in documented, importable formats. The signed, portable export already ships today; the rest is the work.
• Horizon (Phase 11): federation — reach that extends to the people you choose — plus a Chief of Staff on your iPhone drawing on your Macs’ inference, and Signals triage: a layer that watches your tools and events and raises only what genuinely needs you. Its signal primitives already ship in the pack spec today.

Which is the honest note to end a first post on: this is a Developer Preview. Some of the best parts are shipped; some of the best parts are still ahead. What won’t change is the part above the roadmap — the terms.

---

Stallari is a local-first AI platform for macOS, currently in Developer Preview — join the waitlist.