What are AgentScripts?

AgentScripts are user-built WebMCP tools. You install one, and the AI agent in your IDE, your autonomous desktop runner, or your browser gains a set of named, typed tools on a site you already use. When the agent calls a tool, the code runs inside your authenticated browser session.

Yours to write, acts as you, asked before it acts

Three properties define an AgentScript. Yours to write: you describe what the tool does in plain language and Customaise generates the script. Acts as you: the code runs inside your session, with your cookies, your 2FA already cleared, your session keys. Asked before it acts: every tool ships with an allow / prompt / deny consent setting, and the prompt-level gate fires through the Customaise UI which the agent cannot forge.

Why agents need user-built tools

In 2026, the agent is no longer in the IDE alone. Cursor drafts your code. Claude Code ships it. ChatGPT Agent schedules it. OpenClaw runs your workflows while you sleep. Atlas, Comet, and Dia put the agent inside the browser. None of them ship with tools for the sites you actually use. You add them, you grant consent, you stay in control.

What AgentScripts are built on

• UserScripts: the @match, @grant, @require directives and the GM_* API. A mature, battle-tested foundation for running privileged JavaScript on specific sites.
• Model Context Protocol (2024): typed tool discovery and invocation. The same protocol that powers Claude Code's tool calls and Cursor's MCP integrations.
• Human-in-the-loop consent: allow, prompt, deny per tool. The prompt-level gate fires through the Customaise UI which the agent cannot forge.

Cheaper, faster, deterministic

A few hundred tokens per call versus the kilobytes of vision tokens a computer-use agent burns. Milliseconds, not seconds, because the tool is a function call instead of a screenshot-and-OCR loop. Same input, same output: AgentScripts are deterministic; vision-LLM agents are probabilistic.