Hearth is a local-first coding agent that runs entirely on open-source models — no Anthropic or OpenAI key, no account, nothing leaves your machine. We're not chasing a bigger model; we're building the harness to bring open models to Claude-Code-level capability, on hardware you already own. It isn't shipped yet — here's exactly what's real.
Every AI coding tool today works the same way underneath: your code leaves your machine, travels to somebody else's server, and comes back through an API you don't control and can't inspect. That's a rented fire. It's warm while you're paying for it, and it goes cold the moment you stop.
Hearth is the opposite bet, made two ways at once. First: your code and your prompts never leave your computer — not because we promise to guard them somewhere else, but because there is no somewhere else. Second, the harder one: we think open-source models can reach Claude-Code-level capability without a frontier API behind them — not by finding a bigger model, but by building a harness disciplined enough to keep a smaller one honest.
Frontier labs reach Claude-Code-level capability by throwing more parameters and more compute at the problem. We're testing a different bet: that most of what makes a coding agent trustworthy isn't the model — it's the harness around it. Hearth runs on open-source models only, through Ollama or LM Studio, and narrows the gap with deterministic scaffolding: code holds the state of your task, the model's only job is to pick the next action, and verifier gates check that action before it runs.
Deterministic scaffolding doesn't make the model smarter. It turns silently-wrong answers into honest failures — the kind you can see and fix, instead of the kind that ship.
That's not a finished-product claim. It's the architecture we're betting on — and it's why we think open-source-only is a capability strategy, not just a privacy one.
Some code was never going to leave the building. Regulated codebases. Client contracts that forbid it. A founder who just doesn't want a third party holding the keys to the product. If you've ever closed the tab on an AI coding tool because the terms of service made your stomach drop — you're who we built this for. There's no vendor in the loop to trust.
Local is the default, not a setting buried in a menu. Every shell command runs sandboxed with macOS Seatbelt: confined to your workspace, no network access, and your SSH keys and credentials blocked outright. Before Hearth writes a file or runs a command, you see a plain-English approval card — rendered from structured fields, never from model-generated prose. You're not trusting a summary of what it wants to do. You're looking at the actual thing.
We're a waitlist because Hearth isn't finished, not because it isn't real. Most of what's below already runs in our own testing today — mechanisms, not promises; where something's still being wired in, we say so. Launch it with one command and a warm local UI opens; no cloud dashboard, no login.
Before Hearth writes a file or runs a command, it shows an approval card — Allow once, Always allow, or Don't allow. The card is assembled by Hearth itself, not written by the model, so what you approve is exactly what will happen.
Every shell command runs inside a macOS Seatbelt sandbox confined to your workspace. No network access, and protected files like SSH keys and credentials are blocked outright — even if the model asks for them.
Hearth runs on models served through Ollama or LM Studio, with an optional Hugging Face cloud endpoint if you want it. No Anthropic or OpenAI key, ever — and nothing to sign up for.
The model doesn't hold the state of your task — code does. Its only job is to pick the next action, and a verifier checks the result before it counts as done. Measured finding: this turns silent mistakes into visible, fixable ones.
Hearth speaks MCP: the surface for adding your own tools and servers is built into the UI today. Lighting up live connections — plus skills and subagents — is what we're wiring in next. Extensible by you, never a walled platform.
A local web interface with a live activity feed and a gently breathing ember approval card, launched with a single command. No cloud dashboard, no login screen — you watch the agent work instead of trusting it blindly.
We test against real local models, not numbers on a leaderboard. Model choice is yours — swap what's underneath as better open weights ship, without waiting on us to catch up. This list will change as the harness and the models both improve, and when it does, we'll say so here rather than bury it.
Hearth hasn't shipped, and we're not rounding up. We'd rather tell you exactly what's built than oversell what's coming — and earn your trust slowly instead of spending it on a demo.
Leave your email and we'll send you exactly one message: the day Hearth is ready to run on your machine. No spam, no drip campaign — just the one email when we launch, same as everything else we build.
No account, no sign-up — Hearth runs on your machine.
A few things people ask before joining a waitlist for something unshipped.