cherry-studio/src/main/mcpServers/browser/tools/execute.ts
LiuVaayne d41229c69b
Add browser CDP MCP server with session management (#11844)
*  feat: add CDP browser MCP server

* ♻️ refactor: add navigation timeout for browser cdp

* 🐛 fix: reuse window for execute and add debugger logging

*  feat: add show option and multiline execute for browser cdp

*  feat: support multiple sessions for browser cdp

* ♻️ refactor: add LRU and idle cleanup for browser cdp sessions

* Refactor browser-cdp for readability and set Firefox UA

* 🐛 fix: type electron mock for cdp tests

* ♻️ refactor: rename browser_cdp MCP server to browser

Simplify the MCP server name from @cherry/browser-cdp to just browser
for cleaner tool naming in the MCP protocol.

*  feat: add fetch tool to browser MCP server

Add a new `fetch` tool that uses the CDP-controlled browser to fetch URLs
and return content in various formats (html, txt, markdown, json).

Also ignore .conductor folder in biome and eslint configs.

* ♻️ refactor: split browser MCP server into modular folder structure

Reorganize browser.ts (525 lines) into browser/ folder with separate
files for better maintainability. Each tool now has its own file with
schema, definition, and handler.

* ♻️ refactor: use switch statement in browser server request handler

* ♻️ refactor: extract helpers and use handler registry pattern

- Add successResponse/errorResponse helpers in tools/utils.ts
- Add closeWindow helper to consolidate window cleanup logic
- Add ensureDebuggerAttached helper to consolidate debugger setup
- Add toolHandlers map for registry-based handler lookup
- Simplify server.ts to use dynamic handler dispatch

* 🐛 fix: improve browser MCP server robustness

- Add try-catch for JSON.parse in fetch() to handle invalid JSON gracefully
- Add Zod schema validation to reset tool for consistency with other tools
- Fix memory leak in open() by ensuring event listeners cleanup on timeout
- Add JSDoc comments for key methods and classes

* ♻️ refactor: rename browser MCP to @cherry/browser

Follow naming convention of other builtin MCP servers.

* 🌐 i18n: translate pending strings across 8 locales

Translate all "[to be translated]" markers including:
- CDP browser MCP server description (all 8 locales)
- "Extra High" reasoning chain length option (6 locales)
- Git Bash configuration strings (el-gr, ja-jp)
2025-12-16 09:29:30 +08:00

49 lines
1.6 KiB
TypeScript

import * as z from 'zod'
import type { CdpBrowserController } from '../controller'
import { errorResponse, successResponse } from './utils'
export const ExecuteSchema = z.object({
code: z
.string()
.describe(
'JavaScript evaluated via Chrome DevTools Runtime.evaluate. Keep it short; prefer one-line with semicolons for multiple statements.'
),
timeout: z.number().default(5000).describe('Timeout in milliseconds for code execution (default: 5000ms)'),
sessionId: z.string().optional().describe('Session identifier to target a specific page (default: default)')
})
export const executeToolDefinition = {
name: 'execute',
description:
'Run JavaScript in the current page via Runtime.evaluate. Prefer short, single-line snippets; use semicolons for multiple statements.',
inputSchema: {
type: 'object',
properties: {
code: {
type: 'string',
description: 'One-line JS to evaluate in page context'
},
timeout: {
type: 'number',
description: 'Timeout in milliseconds (default 5000)'
},
sessionId: {
type: 'string',
description: 'Session identifier; targets a specific page (default: default)'
}
},
required: ['code']
}
}
export async function handleExecute(controller: CdpBrowserController, args: unknown) {
const { code, timeout, sessionId } = ExecuteSchema.parse(args)
try {
const value = await controller.execute(code, timeout, sessionId ?? 'default')
return successResponse(typeof value === 'string' ? value : JSON.stringify(value))
} catch (error) {
return errorResponse(error as Error)
}
}