In a few weeks of September 2026, personal AI agents went mainstream. Meta launched Muse on September 8 and opened its Muse Connectors program to developers ten days later. OpenAI announced dots at DevDay on September 29: always-on ChatGPT agents, each with its own cloud computer and browser. And Instinct, the assistant you text or call, raised $1 billion to run real-world errands for people.
These agents are good at doing things. But most useful tasks start with knowing something that only lives on a website: which dentists in San Diego take new patients, what a competitor charges today, which jobs were posted this morning. This post explains why that part is hard for any agent, and how Valendata makes it reliable.
The problem: every agent re-learns the web, every time
When an agent needs data from a website, it usually opens a browser and reads the page, step by step, like a person would. That works for a one-off question. It breaks down when the task repeats:
- It is slow and expensive. Every visit is a fresh browsing session, with a model deciding each click.
- It is not structured. You get a paragraph back, not rows you can sort, filter or export.
- It drifts. Ask on Monday and on Friday and you may get different fields, a different number of results, or a summary instead of the list.
- It breaks quietly. When a site changes its layout, the agent may return fewer results and still sound confident.
Personal agents will be asked the same kinds of questions over and over. They need a better way to get web data than browsing from scratch each time.
What Valendata does
Valendata turns a task on a website into a skill: an endpoint that returns clean, structured rows every time it runs. You describe the task, or do it once while Valendata watches, and it saves the way to get the data. After that:
- It runs on demand or on a schedule, and returns the same fields every time.
- It repairs itself. Each run is compared with earlier runs. When a site changes and data stops coming back, the skill works out which step broke, re-reads the live page and learns the new layout. A repair goes live only once it brings the data back, and every version is saved so you can roll back.
- It handles real websites. Sign-ins, two-step codes and pages that load as you scroll are handled in the run, where the site allows it.
- It chains into workflows. Find businesses on Google Maps, find their emails, export them to a sheet, on a schedule.
There is also a public marketplace of ready-made skills. Clone one and it becomes your own copy: when its author improves the original, an update is offered to you, and nothing changes until you take it.
How it fits into Muse, dots and Instinct
All three agents work the same basic way: you give them a goal, connect the apps they need, and they act. Valendata is built to be one of those connected apps: the one that knows how to get data off the web.
Meta Muse. Muse grows through connectors: apps and services it can use on your behalf. Meta's connector guidelines accept "API or MCP documentation" for directory submissions. Valendata has both: a REST API, and a hosted MCP server that exposes your skills and workflows as tools. A Muse that can call a Valendata skill can answer "find me 20 new remote React jobs in Canada" with real listings, not a guess.
OpenAI dots. A dot works from its own computer and connects to apps through ChatGPT. Valendata already connects to ChatGPT as an MCP app (in Developer mode today): add the server, sign in once, and your skills show up as tools. Instead of a dot browsing a site from scratch every morning, it can call a skill that returns yesterday's and today's results in the same shape.
Instinct. Instinct runs errands from a text message, so speed and reliability matter even more: nobody wants to wait while an agent clicks through a site. Instinct has not opened a public connector program yet. When it does, Valendata's API and MCP server are ready. In the meantime, the same skills work for any agent you build yourself.
Working with an assistant today
You don't have to wait for a directory listing. Valendata works now with Claude, ChatGPT, Cursor and any client that supports remote MCP servers:
- Add a custom connector with the address
https://api.valendata.com/mcp. - Sign in to Valendata and choose what the assistant may do. Runs are billed to your account, and you can disconnect at any time.
- Ask in plain words. The assistant can search the marketplace, clone a skill, run it and hand you the rows, right in the chat.
The full setup steps for each assistant are in our docs: Connect your AI assistant.
Why this matters for agents
Agents are getting better at deciding what to do. The bottleneck is getting dependable facts from a web that was built for people, not programs. A self-repairing skill, called through an API or MCP, gives an agent three things it can't get by browsing: the same fields every time, a history to compare against, and a fix when the site changes.
That's the layer Valendata is building: any website, as a reliable API, for any agent.
Start free or browse ready-made skills.
Muse, dots and Instinct are products and trademarks of Meta, OpenAI and Instinct (Spear Street Technology). Valendata is not affiliated with or endorsed by them; their names and logos are used here only to identify their products.
