Works in local testing. Breaks in production with real clients.
Aliax is the production-grade reliability layer between any vision-language model and a real browser. We don't replace your agent — we give it The Steady Hand, The Flight Recorder, and The Closed Loop.
Coordinates rot in 90 seconds
VLMs return pixel coordinates. The DOM re-renders. The button moves 14px. Your agent clicks the cookie banner. Multiply by a million runs. That's the entire problem space.
Stop your VLM from guessing coordinates.
parseUi() injects a deterministic mapper into the page, stamps every interactive element with a stable hash, returns a Set-of-Mark image to your model. execute() resolves the stamp back to a live node — even after re-renders, even inside React portals.
Visual agents lack Sentry. Aliax is the fix.
When 3× DOM state-hashes match or actions cycle, the Flight Recorder captures the failed state and queues it to the Aliax Inbox — your agentic Sentry. A trained annotator resolves it in 78s median. $0.20 per accepted label. The label retrains the in-context router so the same dead-end doesn't repeat.
Not a managed agent. Not a no-code workflow.
We don't run your agent for you. We don't wrap an LLM you can't swap. We don't sell seats. We ship an SDK, a worker, and a usage meter — execution layer for engineers, not a no-code workflow for operators.
how we build
- 01Every claim is a number.
99.9% target uptime. 78s median label. $0.001 per parse. We publish the math, the worked example, and the status page. No vibes.
- 02Failures are typed.
execute() returns
{ ok, action, execution_tier, coords?, error? }— never raises on browser noise. Stagnation is data, not an exception. Your retry logic stays boring. - 03The SDK is the product.
Three methods. Public source. Public CHANGELOG. If a method's behavior changes, the version bumps and the release note tells you exactly which call site to audit.
- 04Failures become your proprietary training moat.
Every failure captured by the Flight Recorder flows through the annotation network and out as a JSON training dataset. The Closed Loop turns production failures into proprietary fine-tuning data. We get cheaper as the model gets smarter — and we tell you when.
small, opinionated, in-the-codebase.
Aliax is built by engineers who shipped browser automation at scale before VLMs existed. We've watched five generations of locators (xpath → role → semantic → coords → multimodal) and every one of them silently broke in production. This is the layer we wished we'd had.
We don't route you to a sales team. Enterprise contracts go through engineering — whoever's on the call has merged code into the worker that week. We answer Discord. We close issues.