Agent

Testing Phase

The Agent feature is currently still in a testing phase; the following APIs and behavior may change before the official release.

Overview

ScriptCat v1.4 introduces the Agent system, giving user scripts a set of capabilities including AI conversation, browser automation, file management, and scheduled tasks.

Scripts call these capabilities through the CAT.agent.* namespace, and every API requires the corresponding permission to be declared with @grant.

Feature Modules

Module	Permission	Description
Conversation	@grant CAT.agent.conversation	Create AI conversations, send messages, stream responses, define custom tools
DOM Operations	@grant CAT.agent.dom	Page navigation, screenshots, clicking, filling, scrolling, DOM monitoring
Skill	@grant CAT.agent.skills	Install/uninstall/invoke Skill packages
Scheduled Tasks	@grant CAT.agent.task	Cron scheduled tasks, event listening
Model	@grant CAT.agent.model	Query configured model information (read-only)
OPFS Files	@grant CAT.agent.opfs	Read/write Agent workspace files
MCP	@grant CAT.agent.mcp	Manage MCP server connections
Skill Development	—	SKILL.cat.md + SkillScript development guide

Quick Start

The simplest possible Agent script:

// ==UserScript==
// @name        Hello Agent
// @match       *://*/*
// @grant       CAT.agent.conversation
// ==/UserScript==

const conv = await CAT.agent.conversation.create();
const reply = await conv.chat("Hi, please introduce yourself");
console.log(reply.content);

Architecture Overview

The Agent system spans multiple isolated contexts within the browser extension:

User script → Sandbox (isolated execution)
              ↓ WindowMessage
           Offscreen (DOM access)
              ↓ ExtensionMessage
           Service Worker (core scheduling)
              ├── LLM Provider (OpenAI / Anthropic)
              ├── ToolRegistry (tool registration and execution)
              ├── SkillScriptExecutor (Skill script execution)
              ├── MCPClient (MCP protocol client)
              └── TaskScheduler (scheduled task scheduling)

Storage Structure

The Agent stores data using the browser's OPFS (Origin Private File System):

agents/
├── conversations/       # conversation history
├── attachments/         # attachments (images, files)
├── skills/{name}/       # Skill package files
│   ├── SKILL.cat.md
│   ├── scripts/
│   └── references/
├── tasks/               # scheduled task config and execution records
└── workspace/           # user workspace files (the directory opfs_* tools operate on)

Supported Models

Provider	Format	Features
OpenAI-compatible	OpenAI Chat Completions API	Supports GPT-4o, DeepSeek, and other compatible models
Anthropic	Anthropic Messages API	Supports the Claude family, Prompt Caching
Zhipu	Zhipu API	Supports the GLM model family

Add a Provider and API Key under "Model Configuration" in the dashboard to use it.

The Skill Ecosystem

A Skill is a package combining prompts + tool scripts + reference material, letting you inject domain-specific knowledge and custom tools into the Agent.

Official Skill repository: scriptscat/skills

Includes ready-to-use Skills for browser automation, scheduled tasks, a Skill-creation tool, conversation/DOM/config examples, and more.

Installation methods:

• URL install — open the SKILL.cat.md URL directly in the browser; ScriptCat automatically intercepts it and shows the install page. You can also paste the URL under the dashboard's Agent → Skill Management.
• Script install — install programmatically via the CAT.agent.skills.install() API

Checking for updates:

A Skill installed via URL records its install source; the dashboard lets you check for updates and upgrade with one click (based on semver comparison of the version field).

See Skill Management API and Skill Development Guide for details.