cherry-studio/packages/aiCore
LiuVaayne 6d15b0dfd1
feat(mcp): add MCP Hub server for multi-server tool orchestration (#12192)
* feat(mcp): add hub server type definitions

- Add 'hub' to BuiltinMCPServerNames enum as '@cherry/hub'
- Create GeneratedTool, SearchQuery, ExecInput, ExecOutput types
- Add ExecutionContext and ConsoleMethods interfaces

Amp-Thread-ID: https://ampcode.com/threads/T-019b4e7d-86a3-770d-82f8-9e646e7e597e
Co-authored-by: Amp <amp@ampcode.com>

* feat(mcp): implement hub server core components

- generator.ts: Convert MCP tools to JS functions with JSDoc
- tool-registry.ts: In-memory cache with 10-min TTL
- search.ts: Comma-separated keyword search with ranking
- runtime.ts: Code execution with parallel/settle/console helpers

Amp-Thread-ID: https://ampcode.com/threads/T-019b4e7d-86a3-770d-82f8-9e646e7e597e
Co-authored-by: Amp <amp@ampcode.com>

* feat(mcp): integrate hub server with MCP infrastructure

- Create HubServer class with search/exec tools
- Implement mcp-bridge for calling tools via MCPService
- Register hub server in factory with dependency injection
- Initialize hub dependencies in MCPService constructor
- Add hub server description label for i18n

Amp-Thread-ID: https://ampcode.com/threads/T-019b4e7d-86a3-770d-82f8-9e646e7e597e
Co-authored-by: Amp <amp@ampcode.com>

* test(mcp): add unit tests for hub server

- generator.test.ts: Test schema conversion and JSDoc generation
- search.test.ts: Test keyword matching, ranking, and limits
- runtime.test.ts: Test code execution, helpers, and error handling

Amp-Thread-ID: https://ampcode.com/threads/T-019b4e7d-86a3-770d-82f8-9e646e7e597e
Co-authored-by: Amp <amp@ampcode.com>

* docs(mcp): add hub server documentation

- Document search/exec tool usage and parameters
- Explain configuration and caching behavior
- Include architecture diagram and file structure

Amp-Thread-ID: https://ampcode.com/threads/T-019b4e7d-86a3-770d-82f8-9e646e7e597e
Co-authored-by: Amp <amp@ampcode.com>

* ♻️ refactor(hub): simplify dependency injection for HubServer

- Remove HubServerDependencies interface and setHubServerDependencies from factory
- Add initHubBridge() to mcp-bridge for direct initialization
- Make HubServer constructor parameterless (uses pre-initialized bridge)
- MCPService now calls initHubBridge() directly instead of factory setter
- Add integration tests for full search → exec flow

* 📝 docs(hub): add comments explaining why hub is not in builtin list

- Add JSDoc to HubServer class explaining its purpose and design
- Add comment to builtinMCPServers explaining hub exclusion
- Hub is a meta-server for LLM code mode, auto-enabled internally

*  feat: add available tools section to HUB_MODE_SYSTEM_PROMPT

- Add shared utility for generating MCP tool function names (serverName_toolName format)
- Update hub server to use consistent function naming across search, exec and prompt
- Add fetchAllActiveServerTools to ApiService for renderer process
- Update parameterBuilder to include available tools in auto/hub mode prompt
- Use CacheService for 1-minute tools caching in hub server
- Remove ToolRegistry in favor of direct fetching with caching
- Update search ranking to include server name matching
- Fix tests to use new naming format

Amp-Thread-ID: https://ampcode.com/threads/T-019b6971-d5c9-7719-9245-a89390078647
Co-authored-by: Amp <amp@ampcode.com>

* ♻️ refactor: consolidate MCP tool name utilities into shared module

- Merge buildFunctionCallToolName from src/main/utils/mcp.ts into packages/shared/mcp.ts
- Create unified buildMcpToolName base function with options for prefix, delimiter, maxLength, existingNames
- Fix toCamelCase to normalize uppercase snake case (MY_SERVER → myServer)
- Fix maxLength + existingNames interaction to respect length limit when adding collision suffix
- Add comprehensive JSDoc documentation
- Update tests and hub.test.ts for new lowercase normalization behavior

*  feat: isolate hub exec worker and filter disabled tools

* 🐛 fix: inline hub worker source

* 🐛 fix: sync hub tool cache and map

* Update import path for buildFunctionCallToolName in BaseService

*  feat: refine hub mode system prompt

* 🐛 fix: propagate hub tool errors

* 📝 docs: clarify hub exec return

*  feat(hub): improve prompts and tool descriptions for better LLM success rate

- Rewrite HUB_MODE_SYSTEM_PROMPT_BASE with Critical Rules section
- Add Common Mistakes to Avoid section with examples
- Update exec tool description with IMPORTANT return requirement
- Improve search tool description clarity
- Simplify generator output with return reminder in header
- Add per-field @param JSDoc with required/optional markers

Fixes issue where LLMs forgot to return values from exec code

* ♻️ refactor(hub): return empty string when no tools available

*  feat(hub): add dedicated AUTO_MODE_SYSTEM_PROMPT for auto mode

- Create self-contained prompt teaching XML tool_use format
- Only shows search/exec tools (no generic examples)
- Add complete workflow example with common mistakes
- Update parameterBuilder to use getAutoModeSystemPrompt()
- User prompt comes first, then auto mode instructions
- Skip hub prompt when no tools available

* ♻️ refactor: move hub prompts to dedicated prompts-code-mode.ts

- Create src/renderer/src/config/prompts-code-mode.ts
- Move HUB_MODE_SYSTEM_PROMPT_BASE and AUTO_MODE_SYSTEM_PROMPT_BASE
- Move getHubModeSystemPrompt() and getAutoModeSystemPrompt()
- Extract shared buildToolsSection() helper
- Update parameterBuilder.ts import

* ♻️ refactor: add mcpMode support to promptToolUsePlugin

- Add mcpMode parameter to PromptToolUseConfig and defaultBuildSystemPrompt
- Pass mcpMode through middleware config to plugin builder
- Consolidate getAutoModeSystemPrompt into getHubModeSystemPrompt
- Update parameterBuilder to use getHubModeSystemPrompt

* ♻️ refactor: move getHubModeSystemPrompt to shared package

- Create @cherrystudio/shared workspace package with exports
- Move getHubModeSystemPrompt and ToolInfo to packages/shared/prompts
- Add @cherrystudio/shared dependency to @cherrystudio/ai-core
- Update promptToolUsePlugin to import from shared package
- Update renderer prompts-code-mode.ts to re-export from shared
- Add toolSetToToolInfoArray converter for type compatibility

* Revert "♻️ refactor: move getHubModeSystemPrompt to shared package"

This reverts commit 894b2fd487.

* Remove duplicate Tool Use Examples header from system prompt

* fix: add handleModeChange call in MCPToolsButton for manual mode activation

* style: update AssistantMCPSettings to use min-height instead of overflow for better layout control

* feat(i18n): add MCP server modes and truncate messages in multiple languages

- Introduced new "mode" options for MCP servers: auto, disabled, and manual with corresponding descriptions and labels.
- Added translations for "base64DataTruncated" and "truncated" messages across various language files.
- Enhanced user experience by providing clearer feedback on data truncation.

* Normalize tool names for search and exec in parser

* Clarify tool usage rules in code mode prompts and examples

* Clarify code execution instructions and update example usage

* refactor: simplify JSDoc description handling by removing unnecessary truncation

* refactor: optimize listAllActiveServerTools method to use Promise.allSettled for improved error handling and performance

---------

Co-authored-by: Amp <amp@ampcode.com>
Co-authored-by: kangfenmao <kangfenmao@qq.com>
2026-01-07 16:35:51 +08:00
..
src feat(mcp): add MCP Hub server for multi-server tool orchestration (#12192) 2026-01-07 16:35:51 +08:00
AI_SDK_ARCHITECTURE.md Feat/aisdk package (#7404) 2025-09-04 14:03:04 +08:00
package.json refactor: switch yarn to pnpm (#12260) 2026-01-05 22:16:34 +08:00
README.md bump ai core version (#11363) 2025-11-19 18:13:33 +08:00
setupVitest.ts Feat/aisdk package (#7404) 2025-09-04 14:03:04 +08:00
tsconfig.json feat: use biome to format files (#10170) 2025-09-15 19:21:15 +08:00
tsdown.config.ts Feat/aisdk package (#7404) 2025-09-04 14:03:04 +08:00
vitest.config.ts fix: preserve openrouter reasoning with web search (#11505) 2025-11-28 13:56:46 +08:00

@cherrystudio/ai-core

Cherry Studio AI Core 是一个基于 Vercel AI SDK 的统一 AI Provider 接口包,为 AI 应用提供强大的抽象层和插件化架构。

核心亮点

🏗️ 优雅的架构设计

  • 简化分层models(模型层)→ runtime(运行时层),清晰的职责分离
  • 函数式优先:避免过度抽象,提供简洁直观的 API
  • 类型安全:完整的 TypeScript 支持,直接复用 AI SDK 类型系统
  • 最小包装:直接使用 AI SDK 的接口,避免重复定义和性能损耗

🔌 强大的插件系统

  • 生命周期钩子:支持请求全生命周期的扩展点
  • 流转换支持:基于 AI SDK 的 experimental_transform 实现流处理
  • 插件分类First、Sequential、Parallel 三种钩子类型,满足不同场景
  • 内置插件webSearch、logging、toolUse 等开箱即用的功能

🌐 统一多 Provider 接口

  • 扩展注册:支持自定义 Provider 注册,无限扩展能力
  • 配置统一:统一的配置接口,简化多 Provider 管理

🚀 多种使用方式

  • 函数式调用:适合简单场景的直接函数调用
  • 执行器实例:适合复杂场景的可复用执行器
  • 静态工厂:便捷的静态创建方法
  • 原生兼容:完全兼容 AI SDK 原生 Provider Registry

🔮 面向未来

  • Agent 就绪:为 OpenAI Agents SDK 集成预留架构空间
  • 模块化设计:独立包结构,支持跨项目复用
  • 渐进式迁移:可以逐步从现有 AI SDK 代码迁移

特性

  • 🚀 统一的 AI Provider 接口
  • 🔄 动态导入支持
  • 🛠️ TypeScript 支持
  • 📦 强大的插件系统
  • 🌍 内置webSearch(Openai,Google,Anthropic,xAI)
  • 🎯 多种使用模式(函数式/实例式/静态工厂)
  • 🔌 可扩展的 Provider 注册系统
  • 🧩 完整的中间件支持
  • 📊 插件统计和调试功能

支持的 Providers

基于 AI SDK 官方支持的 providers

核心 Providers内置支持:

  • OpenAI
  • Anthropic
  • Google Generative AI
  • OpenAI-Compatible
  • xAI (Grok)
  • Azure OpenAI
  • DeepSeek

扩展 Providers通过注册API支持:

  • Google Vertex AI
  • ...
  • 自定义 Provider

安装

npm install @cherrystudio/ai-core ai @ai-sdk/google @ai-sdk/openai

React Native

如果你在 React Native 项目中使用此包,需要在 metro.config.js 中添加以下配置:

// metro.config.js
const { getDefaultConfig } = require('expo/metro-config')

const config = getDefaultConfig(__dirname)

// 添加对 @cherrystudio/ai-core 的支持
config.resolver.resolverMainFields = ['react-native', 'browser', 'main']
config.resolver.platforms = ['ios', 'android', 'native', 'web']

module.exports = config

还需要安装你要使用的 AI SDK provider:

npm install @ai-sdk/openai @ai-sdk/anthropic @ai-sdk/google

使用示例

基础用法

import { AiCore } from '@cherrystudio/ai-core'

// 创建 OpenAI executor
const executor = AiCore.create('openai', {
  apiKey: 'your-api-key'
})

// 流式生成
const result = await executor.streamText('gpt-4', {
  messages: [{ role: 'user', content: 'Hello!' }]
})

// 非流式生成
const response = await executor.generateText('gpt-4', {
  messages: [{ role: 'user', content: 'Hello!' }]
})

便捷函数

import { createOpenAIExecutor } from '@cherrystudio/ai-core'

// 快速创建 OpenAI executor
const executor = createOpenAIExecutor({
  apiKey: 'your-api-key'
})

// 使用 executor
const result = await executor.streamText('gpt-4', {
  messages: [{ role: 'user', content: 'Hello!' }]
})

多 Provider 支持

import { AiCore } from '@cherrystudio/ai-core'

// 支持多种 AI providers
const openaiExecutor = AiCore.create('openai', { apiKey: 'openai-key' })
const anthropicExecutor = AiCore.create('anthropic', { apiKey: 'anthropic-key' })
const googleExecutor = AiCore.create('google', { apiKey: 'google-key' })
const xaiExecutor = AiCore.create('xai', { apiKey: 'xai-key' })

扩展 Provider 注册

对于非内置的 providers可以通过注册 API 扩展支持:

import { registerProvider, AiCore } from '@cherrystudio/ai-core'

// 方式一:导入并注册第三方 provider
import { createGroq } from '@ai-sdk/groq'

registerProvider({
  id: 'groq',
  name: 'Groq',
  creator: createGroq,
  supportsImageGeneration: false
})

// 现在可以使用 Groq
const groqExecutor = AiCore.create('groq', { apiKey: 'groq-key' })

// 方式二:动态导入方式注册
registerProvider({
  id: 'mistral',
  name: 'Mistral AI',
  import: () => import('@ai-sdk/mistral'),
  creatorFunctionName: 'createMistral'
})

const mistralExecutor = AiCore.create('mistral', { apiKey: 'mistral-key' })

🔌 插件系统

AI Core 提供了强大的插件系统,支持请求全生命周期的扩展。

内置插件

webSearchPlugin - 网络搜索插件

为不同 AI Provider 提供统一的网络搜索能力:

import { webSearchPlugin } from '@cherrystudio/ai-core/built-in/plugins'

const executor = AiCore.create('openai', { apiKey: 'your-key' }, [
  webSearchPlugin({
    openai: {
      /* OpenAI 搜索配置 */
    },
    anthropic: { maxUses: 5 },
    google: {
      /* Google 搜索配置 */
    },
    xai: {
      mode: 'on',
      returnCitations: true,
      maxSearchResults: 5,
      sources: [{ type: 'web' }, { type: 'x' }, { type: 'news' }]
    }
  })
])

loggingPlugin - 日志插件

提供详细的请求日志记录:

import { createLoggingPlugin } from '@cherrystudio/ai-core/built-in/plugins'

const executor = AiCore.create('openai', { apiKey: 'your-key' }, [
  createLoggingPlugin({
    logLevel: 'info',
    includeParams: true,
    includeResult: false
  })
])

promptToolUsePlugin - 提示工具使用插件

为不支持原生 Function Call 的模型提供 prompt 方式的工具调用:

import { createPromptToolUsePlugin } from '@cherrystudio/ai-core/built-in/plugins'

// 对于不支持 function call 的模型
const executor = AiCore.create(
  'providerId',
  {
    apiKey: 'your-key',
    baseURL: 'https://your-model-endpoint'
  },
  [
    createPromptToolUsePlugin({
      enabled: true,
      // 可选:自定义系统提示符构建
      buildSystemPrompt: (userPrompt, tools) => {
        return `${userPrompt}\n\nAvailable tools: ${Object.keys(tools).join(', ')}`
      }
    })
  ]
)

自定义插件

创建自定义插件非常简单:

import { definePlugin } from '@cherrystudio/ai-core'

const customPlugin = definePlugin({
  name: 'custom-plugin',
  enforce: 'pre', // 'pre' | 'post' | undefined

  // 在请求开始时记录日志
  onRequestStart: async (context) => {
    console.log(`Starting request for model: ${context.modelId}`)
  },

  // 转换请求参数
  transformParams: async (params, context) => {
    // 添加自定义系统消息
    if (params.messages) {
      params.messages.unshift({
        role: 'system',
        content: 'You are a helpful assistant.'
      })
    }
    return params
  },

  // 处理响应结果
  transformResult: async (result, context) => {
    // 添加元数据
    if (result.text) {
      result.metadata = {
        processedAt: new Date().toISOString(),
        modelId: context.modelId
      }
    }
    return result
  }
})

// 使用自定义插件
const executor = AiCore.create('openai', { apiKey: 'your-key' }, [customPlugin])

使用 AI SDK 原生 Provider 注册表

https://ai-sdk.dev/docs/reference/ai-sdk-core/provider-registry

除了使用内建的 provider 管理,你还可以使用 AI SDK 原生的 createProviderRegistry 来构建自己的 provider 注册表。

基本用法示例

import { createClient } from '@cherrystudio/ai-core'
import { createProviderRegistry } from 'ai'
import { createOpenAI } from '@ai-sdk/openai'
import { anthropic } from '@ai-sdk/anthropic'

// 1. 创建 AI SDK 原生注册表
export const registry = createProviderRegistry({
  // register provider with prefix and default setup:
  anthropic,

  // register provider with prefix and custom setup:
  openai: createOpenAI({
    apiKey: process.env.OPENAI_API_KEY
  })
})

// 2. 创建client,'openai'可以传空或者传providerId(内建的provider)
const client = PluginEnabledAiClient.create('openai', {
  apiKey: process.env.OPENAI_API_KEY
})

// 3. 方式1使用内建逻辑传统方式
const result1 = await client.streamText('gpt-4', {
  messages: [{ role: 'user', content: 'Hello with built-in logic!' }]
})

// 4. 方式2使用自定义注册表灵活方式
const result2 = await client.streamText({
  model: registry.languageModel('openai:gpt-4'),
  messages: [{ role: 'user', content: 'Hello with custom registry!' }]
})

// 5. 支持的重载方法
await client.generateObject({
  model: registry.languageModel('openai:gpt-4'),
  schema: z.object({ name: z.string() }),
  messages: [{ role: 'user', content: 'Generate a user' }]
})

await client.streamObject({
  model: registry.languageModel('anthropic:claude-3-opus-20240229'),
  schema: z.object({ items: z.array(z.string()) }),
  messages: [{ role: 'user', content: 'Generate a list' }]
})

与插件系统配合使用

更强大的是,你还可以将自定义注册表与 Cherry Studio 的插件系统结合使用:

import { PluginEnabledAiClient } from '@cherrystudio/ai-core'
import { createProviderRegistry } from 'ai'
import { createOpenAI } from '@ai-sdk/openai'
import { anthropic } from '@ai-sdk/anthropic'

// 1. 创建带插件的客户端
const client = PluginEnabledAiClient.create(
  'openai',
  {
    apiKey: process.env.OPENAI_API_KEY
  },
  [LoggingPlugin, RetryPlugin]
)

// 2. 创建自定义注册表
const registry = createProviderRegistry({
  openai: createOpenAI({ apiKey: process.env.OPENAI_API_KEY }),
  anthropic: anthropic({ apiKey: process.env.ANTHROPIC_API_KEY })
})

// 3. 方式1使用内建逻辑 + 完整插件系统
await client.streamText('gpt-4', {
  messages: [{ role: 'user', content: 'Hello with plugins!' }]
})

// 4. 方式2使用自定义注册表 + 有限插件支持
await client.streamText({
  model: registry.languageModel('anthropic:claude-3-opus-20240229'),
  messages: [{ role: 'user', content: 'Hello from Claude!' }]
})

// 5. 支持的方法
await client.generateObject({
  model: registry.languageModel('openai:gpt-4'),
  schema: z.object({ name: z.string() }),
  messages: [{ role: 'user', content: 'Generate a user' }]
})

await client.streamObject({
  model: registry.languageModel('openai:gpt-4'),
  schema: z.object({ items: z.array(z.string()) }),
  messages: [{ role: 'user', content: 'Generate a list' }]
})

混合使用的优势

  • 灵活性:可以根据需要选择使用内建逻辑或自定义注册表
  • 兼容性:完全兼容 AI SDK 的 createProviderRegistry API
  • 渐进式:可以逐步迁移现有代码,无需一次性重构
  • 插件支持:自定义注册表仍可享受插件系统的部分功能
  • 最佳实践:结合两种方式的优点,既有动态加载的性能优势,又有统一注册表的便利性

📚 相关资源

未来版本

  • 🔮 多 Agent 编排
  • 🔮 可视化插件配置
  • 🔮 实时监控和分析
  • 🔮 云端插件同步

📄 License

MIT License - 详见 LICENSE 文件


Cherry Studio AI Core - 让 AI 开发更简单、更强大、更灵活 🚀