An agentic browser harness: an AI side panel & orb in your own logged-in Chrome, Edge, Arc, Brave or even Firefox run by the agent/subscription you already use (Claude Code, Codex, Antigravity). Free, open source, no API key. Also an MCP server.
TypeScript agentic-browseragentic-browsingagy-cliaiai-agentai-browser-extensionai-toolsbrave-extensionbrowser-agentbrowser-automationbrowser-extensionbrowser-usebrowserskillchrome-extensionclaude-codecodexfirefox-addonmcp-servernanobrowserweb-automation
import { z } from 'zod';
import { ActionError, defineAction } from '../core';
import { digestOf, groupEnd, readCursor, writeCursor } from './chunks';
import { cssPath, resolveTarget, targetSchema } from './dom';
const MAX_EXTRACT_CHARS = 200_000;
export const extractText = defineAction({
name: 'page.extractText',
description:
'Read the rendered text or raw HTML of an element or the whole page, one group at a time. ' +
'A group is cut at the last paragraph or sentence boundary that fits in maxLength, never mid-sentence, ' +
'and the same page always groups the same way. When more text remains the reply carries a cursor — pass it ' +
'back for the next group, and keep going until no cursor comes back. ' +
'The cursor is checked against the text it already handed you, so two groups never come from two different ' +
'versions of the page: if the page rewrote what you had read, the reply is "stale": true and nothing else, ' +
'and the read starts over with no cursor.',
input: z.object({
target: targetSchema.optional().describe('Element to read; defaults to the whole page'),
format: z
.enum(['text', 'html'])
.default('text')
.describe(
'Plain rendered text, or raw HTML. HTML is denied by default — it carries comments and hidden nodes ' +
'the reader never sees — and needs "raw-html-read": "allow" in the guardrail config.',
),
maxLength: z
.number()
.int()
.positive()
.max(MAX_EXTRACT_CHARS)
.default(20000)
.describe('Characters per group — the reply stops at the last boundary that fits, so it comes back a little shorter'),
cursor: z
.string()
.max(32)
.optional()
.describe('Continue a read: the cursor the previous reply returned. Omit to start from the top.'),
}),
execute({ target, format, maxLength, cursor }) {
const element = target ? resolveTarget(target) : document.body;
const full =
format === 'text' ? element.innerText.replace(/\n{3,}/g, '\n\n').trim() : element.outerHTML;
const source = cssPath(element);
const budget = Math.min(maxLength, MAX_EXTRACT_CHARS);
const resume = cursor ? readCursor(cursor) : null;
if (cursor && !resume) {
throw new ActionError(
'That is not a cursor page.extractText handed out — omit it to read from the top',
'INVALID_INPUT',
);
}
if (resume && digestOf(full.slice(0, resume.offset)) !== resume.digest) {
return { stale: true, source, length: full.length };
}
const offset = resume?.offset ?? 0;
const end = groupEnd(full, offset, budget);
const more = end < full.length;
return {
content: full.slice(offset, end).trimEnd(),
source,
length: full.length,
offset,
truncated: more,
...(more ? { nextOffset: end, cursor: writeCursor(end, full.slice(0, end)) } : {}),
};
},
});