Importing Websites as Skills: Build Composable Browser Agents

​Most of the web has no API. The moment your agent has to operate a real site the way a person would, it runs into the least reliable, least observable part of the whole stack: the browser. And the moment it acts for someone, it's handling data you have to protect, on sites you don't control. ​In this session, Mastra and Browserbase show how the two halves fit together. Browserbase's browse.sh, an open catalog of website skills, lets an agent import any site as a reusable skill, and learn a brand-new one live when the catalog doesn't have it yet. Mastra runs those skills as durable, observable workflows, so every step is recoverable when it fails and auditable after the fact. What to expect:We'll build a relocation concierge live: one agent whose job is to help someone move to a new city. It composes skills across the catalog (neighborhoods, a gym nearby, furniture, the move itself), then does the hard part: finding an apartment and applying for it on the user's behalf. You'll see a real action taken with the user's SSN redacted before it leaves the run, plus a full replay of what the agent did, so a support person can debug a broken step without ever seeing the user's identity. ​Whether you're building agents that have to operate real websites with no API, or keeping them safe and auditable once real user data is in play, you'll leave with a concrete pattern for both.