Hey HN! We're Toni and Pablo, and we're building Karumi (https://www.karumi.ai), a system that lets your users get instant, scalable, guided demos of your product, fully automated, zero human interaction. It works in any language.

There's a live demo at https://www.karumi.ai/meet/start/phlz.

Karumi is an AI agent that operates a real web app in a shared browser session and talks the user through it. Instead of a human giving a screen-share demo, the agent opens your product, clicks around, fills forms, and explains what it’s doing. We started building this as an internal tool at our previous company. As the product grew, people kept asking “what’s the right way to demo feature X?“. Docs and scripts became outdated quickly, and the quality of demos depended too much on who was presenting. We wanted something closer to a repeatable program: an agent that knows the main flows, understands who it’s talking to, and can walk through the product without getting lost.

Over time this turned into three main components:

Planning/control layer

A loop that decides the next step: ask something, click, navigate, reset, etc. It uses a reasoning model, but only within a fixed set of allowed actions with guards (timeouts, depth limits, reset states). It never gets free control of the browser.

Browser execution layer

A controlled browser session, streamed in a video call. The agent can only interact with the elements we want. We log each action with a timestamp and the agent’s “reason”, which helps debug odd behavior.

Product knowledge layer

We ingest docs, demo scripts & videos, and usage analytics, to train the agent. At runtime, the agent uses its knowledge to decide what flow to show and how to explain it.

Some practical details and limitations:

We only support web apps right now. Desktop apps will come next. LLMs introduce non-determinism, so we bias toward safe, predictable behavior: checkpoints, conservative navigation, and “escape hatches” that reset to known states. If the agent doesn’t understand a UI state (unknown modal, layout shift, etc.), it asks the user instead of guessing. Regarding pricing, it’s still early. We tailor it to each customer based on their needs. Our current thinking is a platform fee plus a per-call charge for the agent. The platform fee varies depending on complexity, support requirements, and overall scope.

People currently use Karumi for inbound demos and internal demo environments. If you want to see it inside a real product, here’s Karumi running in Deel’s platform: https://www.loom.com/share/e7f7e00f2284478e8335f8f4d4dac6bd

We’ll be around to answer questions and look forward to your feedback!