Roadmap
The mission, stated plainly: local autonomous computer usage — an AI that does real work on your machine, where nothing you do leaves it. "Run LLMs locally" is the foundation, not the destination. Here's the concrete path, stage by stage, with what must be true before each step ships.
Stage 1 — Private chat, everywhere
The foundation is live: native apps on macOS and iOS (one Swift codebase), Android (Kotlin), downloadable previews for Windows and Linux, and a CLI for programmers (announced today). Under it: a 12-model catalog with RAM-fit gating and SHA-256-verified downloads, a task router that suggests the right installed model per job, repetition-loop and encoding guardrails, and cross-platform logic kept identical by machine-enforced parity tests. Everything measured is published on the reality page; everything broken is in the failure-mode ledger.
Stage 2 — A model that can see your work
Before an AI can act on your computer, it has to perceive what you give it:
- Documents in chat — ✅ shipped (July 2026). Attach a text file or (on Mac/iOS) a PDF, ask about it; follow-up questions keep it in context. Binary files are refused with a reason, never silently mangled.
- Images — the models are ready (Gemma 4 is fully multimodal); our engine builds are text-only today. Requires rebuilding the inference layer with multimodal support on every platform — which is why images don't attach yet: we don't ship buttons before the pipeline behind them works.
- Frictionless install — Microsoft Store, Flathub, winget, an npm package for the CLI; one shared model store so the app and CLI never download the same gigabytes twice.
Stage 3 — The agent grows hands
The first slice shipped in July 2026 (announcement): grant the agent a workspace folder and it can list, read, move, rename, create, and trash-to-a-visible-folder — proposing a plan you approve once, with per-step undo, a plain-text ledger of everything it did and tried, and structural guarantees: it names files but can never mint paths, nothing overwrites, nothing writes without your per-run yes, and a plan with one blocked step is refused whole.
Since then, on macOS, the hands got much longer — all on the same consent/undo/ledger spine. The agent can now click and type in any app through the accessibility layer (buttons, menu bar, navigation keys — not just AppleScript-scriptable apps) and run any of your Apple Shortcuts, so it reaches most of what you can do on the machine. Longer multi-step tasks stream each step live so you watch and can interrupt (Ctrl-C stops instantly), and it recovers gracefully when a small local model gets stuck. Three trust features a cloud agent can't match arrived too: undo a whole task in a later session, a per-task audit trail you can read back, and a real dry run that shows exactly what it would do while changing nothing. What remains in Stage 3: more chore classes and deeper app coverage, the polished desktop approval dialog (the CLI already works end-to-end today), and bringing the same reliability guards to the mobile apps.
Stage 4 — Local autonomous computer usage
The end state: you describe an outcome, and an agent on your machine does the work — sorts the folder, drafts the reply, runs the pipeline — with the same privacy guarantee as chat, because it's the same architecture: no server, no account, nothing leaves the device. That's a capability cloud agents can't offer, and it's the whole reason this project exists.
Why local wins (the honest comparison)
Cloud computer-use agents are smarter per step — we don't pretend otherwise. But for the chores you'd actually hand an agent, a local one wins on what matters: stop it dead mid-task, read everything it did and refused, undo the whole task with one click, and no AI middleman ever sees your files. The full argument, including where we lose: why a local agent beats a cloud one.
How this stays funded (and free)
The chat apps, the model library, and the CLI are free and MIT-licensed — permanently. The plan is open-core: when the autonomy layer (Stages 3–4) matures, that becomes the paid tier for the people who need it, and it funds everything else. No ads, no data sales — there's no data to sell; that's the point.
Help us get there — we're hiring
Stage 4 needs people. We believe local models are the future of AI, and the only thing between here and there is making small, on-device models genuinely capable — a research problem worth a career. We're hiring researchers in on-device inference, small-model agentic capability, and honest evaluation. Read the bet and the open roles →
This page changes as reality does — no dates you can screenshot and shame us with later, just order and honesty. Watch progress commit-by-commit on GitHub.