# Your agent can click through a site. Now it can leave a program behind.

Coding agents are good at exploring a site and bad at doing it the same way twice. With Kaabist over MCP, the agent explores in a real browser, and what it learns becomes a fixed program that runs the same way every time, without the agent.

## The problem

An agent that drives a browser on every run is slow, costs tokens each time, and takes a slightly different path every time. That is fine for exploring, and wrong for anything you run on a schedule.

## Explore with the agent

Connect your agent to mcp.kaab.ist. It opens a Kaabist browser, reads a condensed view of each page, and acts on ranked locators instead of guessing at selectors.

```
claude mcp add --transport http kaab https://mcp.kaab.ist/mcp
```

## Keep what worked

The agent turns its session into a draft program, lints it, and dry-runs it. You review the program, not a transcript.

## Run it without the agent

The published program runs on Kaabist, on a schedule or from your API calls, the same way every time. The agent is only needed again when you want to change it.

## Questions

### Does every run cost model tokens?

No. Runs execute the program deterministically. A language model is optional and only helps when a locator needs healing.