Every SaaS product with an AI feature ends up with the same chart in a spreadsheet somewhere: inference cost climbing in a straight line next to user growth. We built Locai to bend that line down, and this week we shipped the piece that lets you wire it directly into your own product.
Locai:Link, our on-device inference runtime, now integrates into any application. Not as a separate tool your users have to find and install themselves, but as a sidecar that ships inside your own installer, talks to your code over a local, OpenAI-compatible endpoint, and redirects inference to your end users’ own hardware, at scale, orchestrated centrally through Locai:Control.
This is aimed at two kinds of frustration we keep hearing about. One is developers trying to build the next AI-powered product without a cloud bill that scales against them from day one. The other is end users who’ve got understandably twitchy about where their data goes every time an app calls out to a cloud model. If either of those sounds familiar, or if the cost problem is still on the horizon rather than in front of you yet, this is the update worth trying.
What’s actually new here
Until now, Link showed up in the places we’d built direct support for: VS Code, Open WebUI, Obsidian, OpenClaw. If you wanted your own application to push inference onto a user’s device, you were mostly building that plumbing yourself. This release adds “Your Own Software” as a first-class integration path, so Link becomes something you can drop into any product, not just the tools we’ve already wired up.
The mechanics are deliberately boring:
- Your installer bundles Link and starts it alongside your app. Users don’t manage it separately.
- Link exposes an OpenAI-compatible chat completions endpoint on the user’s machine. If your app already calls a hosted model, pointing it at Link is closer to a config change than a rewrite.
- Prompts, responses, and model weights stay on the device. Locai:Control only ever sees operational metadata: device health, agent version, whether a model is currently serving. Not what your users are asking or what they’re getting back.
- Link updates itself in the background, so you’re not shipping new binaries to every user’s machine every time we patch something.
- The same plugin architecture handles more than chat. Audio transcription and image and audio classification run through the same sidecar today, with more modalities coming, so if your roadmap goes beyond text, you’re not building a second integration later.
Why this matters, depending on who you are
If you’re running a SaaS product and your biggest customers are also your biggest cloud AI bill, this is the fix we built the whole platform around: move the inference that doesn’t need to happen in a data centre onto the device that’s already sitting there, idle, most of the day. That’s the actual economics change, not a discount on the same bill, a different bill entirely. We go into the unit economics in more detail on the For SaaS page, but the short version is that this integration is what turns “move inference to the edge” from a slide in a deck into something your engineers can actually ship.
If you’re in a regulated industry, the driver usually isn’t cost, it’s where the data is allowed to go. Link running as a sidecar inside your own application means inference happens on infrastructure you or your customer already controls, with nothing routed through a third party’s servers by default. We cover the compliance side, including air-gapped deployments, on the For Enterprise page.
Getting started
The integration is documented end to end, including the health check your app should run before attempting inference, the CORS setup for browser-based apps, and how Link’s self-update process works so it doesn’t surprise you mid-release. If you’re an engineer, start with the integration docs. It’s a short read, and the quick start gets you from a healthz check to a working chat completion in a handful of lines.
If you’d rather talk it through first, get in touch and we’ll walk through what a custom setup would look like for your product. And one more thing, from me directly: if you’re a Fuel Ventures portfolio company, we’re offering meaningfully discounted rates for the first wave of teams integrating this. Drop me a line and we’ll sort it out.
A few questions we expect to get
Does this replace our existing cloud AI setup? Not usually, at least not on day one. Most teams route the requests that make sense locally, things that are frequent, latency-sensitive, or don’t need a frontier model, and keep the rest on their existing provider. Link and a cloud API can run side by side.
What happens if a user’s device can’t run Link, or it isn’t installed? Your app checks a health endpoint before it tries to run inference locally. If Link isn’t reachable, you fall back however you’d like: prompt the user to install it, or route that request to the cloud instead. Nothing forces an all-or-nothing switch.
Which platforms are supported? Link ships today for macOS (arm64), Linux (x86_64), and Windows (x86_64), distributed as a signed bundle your installer extracts and starts.
Is this only for text models? No. The same sidecar also runs audio transcription and image and audio classification through its plugin system, so a single integration covers more than chat completions if you need it to.