Skip to content
Veilus Docs
These docs describe the Veilus release after v0.2.1, coming soon. If you have v0.2.1, some screens and features (Datasets, Trash, the new profile panel, many MCP tools) are not in your version yet.

AI Agent over MCP

This recipe lets Claude drive a Veilus profile: look at a page, fill a form, check the result, then save what it did as a script you approve and schedule. The example fills an order form once per row of a dataset.

Prompt, script and example data: solutions/ai-agent-mcp on GitHub. script.ts runs daily on a real Veilus profile.

Two ways to use it

Section titled “Two ways to use it”
  • With Claude: needs Claude Code with the Veilus plugin (/plugin marketplace add veilus/claude-plugin, then /plugin install veilus@veilus) and the Veilus app running with the local API on. Paste the prompt from prompt.md. Claude works in the profile through the Veilus MCP server and saves a script. Nothing runs unattended until you approve it in Veilus Flow → Scripts. See also Let an LLM do the automation for you.
  • Without Claude: paste script.ts into Veilus Flow → Scripts and approve it (dataset rows reach approved scripts only), import dataset.csv and assign it to a profile. See Scripts and Datasets.

What a run prints

Section titled “What a run prints”
RESULT {"submitted":2,"orders":[{"name":"Alex Doe","size":"medium"}, …]}

The script throws if the site did not take an order, so a failed submit shows as a failed run.

Change it for your form

Section titled “Change it for your form”
  • Profile variable FORM_URL: your form. Then change the page.fill selectors and the check after the submit. The example reads httpbin’s JSON echo, while a real site shows a confirmation page. Or ask Claude to redo step 4 of the prompt on your page.
  • dataset.csv: one row per submission, columns NAME, PHONE, EMAIL, SIZE, COMMENTS.

The prompt is not tested automatically: an LLM writes different code on each run, so read the script before you approve it.

Section titled “Legal considerations”