MCP 服务器

anygpt-discovery

Exposes 5 meta-tools that allow AI agents to autonomously discover and execute tools from 100+ MCP servers, reducing token usage by 99%.

README

AnyGPT Ecosystem

⚠️ WORK IN PROGRESS: This project is under active development. APIs, components, and configurations may change significantly. Use at your own risk in production environments.

A comprehensive TypeScript ecosystem for building AI-powered applications with support for multiple providers, MCP protocol, and flexible configuration management.

Why?

Problem: Building AI applications requires dealing with different provider APIs, complex configuration management, and protocol translations for MCP clients. Solution: This monorepo provides a modular ecosystem with clean separation of concerns:

Type System: Pure type definitions with zero runtime overhead
Router Layer: Provider abstraction and routing with connector pattern
Configuration: Dynamic connector loading and flexible configuration management
CLI Interface: Command-line tool for AI interactions and conversation management
MCP Server: Protocol translator for MCP clients (Docker Desktop, Windsurf, etc.)

Architecture

┌─────────────────────────────────────────────────────────────────┐
│ MCP Clients (Claude Desktop, Windsurf, Cursor)                 │
└─────────────────────────────────────────────────────────────────┘
                    ↓                           ↓
    ┌───────────────────────────┐   ┌─────────────────────────┐
    │ @anygpt/mcp-discovery-server │   │ @anygpt/mcp           │
    │ (5 meta-tools, 99% token    │   │ (MCP protocol server) │
    │  reduction)                  │   └─────────────────────────┘
    └───────────────────────────┘               ↓
                    ↓                   ┌─────────────────┐
        ┌───────────────────────┐       │ @anygpt/config  │
        │ @anygpt/mcp-discovery │       │ (Configuration) │
        │ (Discovery engine)    │       └─────────────────┘
        └───────────────────────┘               ↓
                                        ┌─────────────────┐
┌─────────────────────────┐             │ @anygpt/router  │
│ @anygpt/cli             │────────────→│ (Provider       │
│ (Command-line tool)     │             │  abstraction)   │
└─────────────────────────┘             └─────────────────┘
                                                ↓
                                    ┌───────────────────────┐
                                    │ @anygpt/ai-provider   │
                                    │ (Function calling)    │
                                    └───────────────────────┘
                                                ↓
                        ┌───────────────────────────────────────┐
                        │ Connectors (@anygpt/openai,           │
                        │  @anygpt/anthropic, @anygpt/cody,     │
                        │  @anygpt/claude, @anygpt/mock)        │
                        └───────────────────────────────────────┘
                                                ↓
                        ┌───────────────────────────────────────┐
                        │ AI Provider APIs (OpenAI, Anthropic,  │
                        │  Ollama, LocalAI, etc.)               │
                        └───────────────────────────────────────┘

Supporting Packages:
  • @anygpt/types - Pure type definitions (0 runtime deps)
  • @anygpt/rules - Type-safe rule engine for configuration
  • @anygpt/mcp-logger - File-based logging for MCP servers
  • @anygpt/plugins - Plugin system (docker-mcp-plugin, etc.)

Core Packages

Package	Purpose	Dependencies
@anygpt/types	Pure type definitions	None (0 runtime deps)
@anygpt/config	Configuration management	@anygpt/types
@anygpt/router	Core routing and connector registry	None
@anygpt/ai-provider	AI provider wrapper with function calling	@anygpt/router
@anygpt/rules	Type-safe rule engine	None
@anygpt/mcp-logger	File-based logging for MCP servers	None
@anygpt/plugins	Plugin system for dynamic configuration	@anygpt/types
@anygpt/mcp-discovery	MCP tool discovery engine (core logic)	@anygpt/types
@anygpt/mcp-discovery-server	MCP Discovery Server (PRIMARY interface)	@anygpt/mcp-discovery
@anygpt/cli	Command-line interface	@anygpt/config
@anygpt/mcp	MCP server implementation	@anygpt/config

Connector Packages

Package	Provider	Dependencies
@anygpt/openai	OpenAI & compatible APIs	@anygpt/router, openai
@anygpt/anthropic	Anthropic Claude (native)	@anygpt/router, @anthropic-ai/sdk
@anygpt/claude	Claude via MCP	@anygpt/router
@anygpt/cody	Sourcegraph Cody	@anygpt/router
@anygpt/mock	Testing & development	@anygpt/types

Supported Providers

OpenAI: GPT-4o, GPT-4, GPT-3.5, o1 models
OpenAI-Compatible: Ollama, LocalAI, Together AI, Anyscale
Anthropic: Claude Sonnet, Opus, Haiku (native API)
Mock Provider: For testing and development

Quick Start

Install CLI Tool

npm install -g @anygpt/cli

Install Individual Packages

# For building applications
npm install @anygpt/router @anygpt/openai @anygpt/anthropic

# For configuration management
npm install @anygpt/config

# For type definitions only
npm install @anygpt/types

MCP Discovery Server (99% Token Reduction!)

Zero-configuration MCP server that enables AI agents to discover and execute tools from 100+ MCP servers without loading everything into context.

# No installation needed - use with npx
npx -y @anygpt/mcp-discovery-server

Add to Claude Desktop / Windsurf / Cursor:

{
  "mcpServers": {
    "anygpt-discovery": {
      "command": "npx",
      "args": ["-y", "@anygpt/mcp-discovery-server"]
    }
  }
}

What it does:

Exposes 5 meta-tools instead of 150+ individual tools
AI agents autonomously discover tools using search_tools
Reduces token usage from 100,000+ to ~600 tokens (99% reduction)
Gateway capability: discover AND execute tools through single connection

Example workflow:

User: "Read README.md and create a GitHub issue if there are TODOs"

AI Agent:
1. search_tools({ query: "read file" }) → finds filesystem:read_file
2. execute_tool({ server: "filesystem", tool: "read_file", ... })
3. search_tools({ query: "create github issue" }) → finds github:create_issue
4. execute_tool({ server: "github", tool: "create_issue", ... })

Result: ~1,000 tokens vs 500,000+ tokens (99.8% savings)

MCP Server

# Install and run MCP server
npm install -g @anygpt/mcp
anygpt-mcp

Usage Examples

1. CLI Usage

AI Chat Commands

# Discover available models and tags
anygpt list-tags

# Quick chat with tags (stateless)
anygpt chat --tag sonnet "Explain TypeScript generics"
anygpt chat --tag opus "Write a complex algorithm"

# Specify provider explicitly
anygpt chat --provider cody --tag sonnet "Hello"
anygpt chat --provider provider1 --tag gemini "Hello"

# Use direct model name (no tag resolution)
anygpt chat --model "ml-asset:static-model/claude-sonnet-4-5" "Hello"

# Start a conversation (stateful)
anygpt conversation start --tag sonnet --name "coding-session"
anygpt conversation message "How do I implement a binary tree in TypeScript?"
anygpt conversation message "Show me the insertion method"

# List conversations
anygpt conversation list

# Fork a conversation with different tag
anygpt conversation fork --tag opus --name "binary-tree-v2"

MCP Management Commands

# List all MCP servers
anygpt mcp list
anygpt mcp list --enabled    # Only enabled servers
anygpt mcp list --disabled   # Only disabled servers

# Search for tools across all servers
anygpt mcp search "github"
anygpt mcp search "create" --server github-official

# Inspect tool details (auto-resolves server)
anygpt mcp inspect search
anygpt mcp inspect create_issue --server github-official

# Execute tools with natural syntax
anygpt mcp execute search "how to cook paella"
anygpt mcp execute search "query" 5  # Multiple parameters

# List tools from specific server
anygpt mcp tools github-official
anygpt mcp tools github-official --all  # Include disabled tools

2. Router as Library

import { GenAIRouter } from '@anygpt/router';
import { OpenAIConnectorFactory } from '@anygpt/openai';

// Create router and register connector
const router = new GenAIRouter();
router.registerConnector(new OpenAIConnectorFactory());

// Create connector instance
const connector = router.createConnector('openai', {
  apiKey: process.env.OPENAI_API_KEY,
  baseURL: 'https://api.openai.com/v1',
});

// Make requests
const response = await connector.chatCompletion({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'Hello!' }],
});

3. Docker MCP Plugin

Auto-discover and configure Docker MCP servers:

// anygpt.config.ts
import { defineConfig } from '@anygpt/config';
import DockerMCP from '@anygpt/docker-mcp-plugin';

export default defineConfig({
  plugins: [
    DockerMCP({
      serverRules: [
        // Disable specific servers
        {
          when: { name: 'sequentialthinking' },
          set: { enabled: false },
        },
      ],
    }),
  ],
});

What it does:

Discovers all Docker MCP servers automatically
Creates separate MCP server instance for each
Supports server-level enable/disable rules
Disabled servers still visible for discovery

Generated configuration:

{
  mcpServers: {
    'github-official': {
      command: 'docker',
      args: ['mcp', 'gateway', 'run', '--servers', 'github-official'],
      source: 'docker-mcp-plugin',
      metadata: { toolCount: 49 }
    },
    'duckduckgo': {
      command: 'docker',
      args: ['mcp', 'gateway', 'run', '--servers', 'duckduckgo'],
      source: 'docker-mcp-plugin',
      metadata: { toolCount: 2 }
    }
  }
}

4. Configuration-Driven Setup

import { setupRouter } from '@anygpt/config';

// Automatically loads config and sets up router
const { router, config } = await setupRouter();

// Use with any registered connector
const response = await router.chatCompletion({
  provider: 'openai-main',
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'Hello!' }],
});

Factory config example (.anygpt/anygpt.config.ts):

import { config } from '@anygpt/config';
import { openai } from '@anygpt/openai';
import { anthropic } from '@anygpt/anthropic';

export default config({
  defaults: {
    provider: 'openai-main',
    model: 'gpt-4o',
  },
  providers: {
    'openai-main': {
      name: 'OpenAI GPT Models',
      connector: openai({
        apiKey: process.env.OPENAI_API_KEY,
        baseURL: 'https://api.openai.com/v1',
      }),
    },
    claude: {
      name: 'Anthropic Claude',
      connector: anthropic({
        apiKey: process.env.ANTHROPIC_API_KEY,
        baseURL: 'https://api.anthropic.com', // Optional: for corporate gateways
      }),
    },
    'ollama-local': {
      name: 'Local Ollama',
      connector: openai({
        baseURL: 'http://localhost:11434/v1',
      }),
    },
  },
});

Alternative standard config format:

import type { AnyGPTConfig } from '@anygpt/config';

const config: AnyGPTConfig = {
  version: '1.0',
  providers: {
    'openai-main': {
      name: 'OpenAI GPT Models',
      connector: {
        connector: '@anygpt/openai',
        config: {
          apiKey: process.env.OPENAI_API_KEY,
          baseURL: 'https://api.openai.com/v1',
        },
      },
    },
  },
  settings: {
    defaultProvider: 'openai-main',
    timeout: 30000,
  },
};

export default config;

5. MCP Server Usage

# Run MCP server
anygpt-mcp

# Test with MCP Inspector
npx @modelcontextprotocol/inspector anygpt-mcp

Claude Desktop Integration:

{
  "mcpServers": {
    "anygpt": {
      "command": "anygpt-mcp",
      "env": {
        "OPENAI_API_KEY": "your-openai-api-key"
      }
    }
  }
}

Development

This project uses NX monorepo for managing multiple packages:

# Install dependencies (automatically installs Husky git hooks)
npm install

# Note: package-lock.json is created locally but never committed
# - Nx requires it for builds
# - Husky pre-commit hook auto-unstages it
# - You can use any npm registry (public or internal)

# Build all packages (NX handles dependencies automatically)
npx nx run-many -t build

# Build specific package (dependencies built automatically)
npx nx build cli

# Run tests
npx nx run-many -t test

# Run E2E tests
npx nx e2e e2e-cli

# Lint all packages
npx nx run-many -t lint

Package Dependency Graph

@anygpt/types (no deps)
    ↓
@anygpt/config, @anygpt/mock
    ↓
@anygpt/router → @anygpt/openai
    ↓
@anygpt/cli, @anygpt/mcp

Key Features

🎯 Modular Architecture

Clean separation: Each package has a single responsibility
Zero runtime overhead: Type-only packages with import type
Dependency inversion: Connectors depend on router, not vice versa

🔧 Dynamic Configuration

Runtime connector loading: No hardcoded dependencies
Multiple config sources: TypeScript, JavaScript, JSON files
Environment support: User home, system-wide, project-local configs
Plugin system: Auto-discovery and configuration generation

🔍 MCP Discovery & Management

Auto-discovery: Automatically finds and configures Docker MCP servers
Tool-focused CLI: Execute tools without knowing which server provides them
Server rules: Enable/disable servers while maintaining visibility
Variadic arguments: Natural command-line syntax for tool execution
On-demand loading: Load only what you need, when you need it

🚀 Developer Experience

Full TypeScript support: Complete type safety across all packages
Comprehensive CLI: Stateful conversations, forking, summarization, MCP management
Testing utilities: Mock connector for development and testing
Progress indicators: Visual feedback for long-running operations

🔌 Extensible Design

Connector pattern: Easy to add new AI providers
Plugin architecture: Extensible command system with auto-discovery
MCP compliance: Full protocol implementation
Separate server instances: Each MCP server runs independently

✅ Comprehensive Testing

30 E2E tests: Complete CLI workflow validation with 0 skipped tests
Mock connector: Deterministic responses for reliable testing
Full coverage: Chat, conversations, config management, and error handling
CI/CD ready: Fast, reliable tests that run in < 15 seconds

Documentation

Getting Started

CLI Documentation - Complete command-line interface guide
Configuration Guide - Complete configuration setup and examples
Product Documentation - Features, architecture, and use cases
Troubleshooting Guide - Common issues, recent fixes, and debugging

Integration Examples

Docker cagent Integration - Use AnyGPT MCP Discovery with Docker cagent for intelligent multi-agent systems (99% token reduction!)
LiteLLM Integration - Use AnyGPT with LiteLLM Proxy for 100+ providers and enterprise features
Example Configurations - Ready-to-use config examples for various setups

Development Guidelines

Testing Guide - Comprehensive testing strategy, patterns, and coverage goals
E2E Testing Guide - End-to-end test suite documentation and patterns
Release Workflow - Automated Release PR workflow documentation
Release Quick Reference - Quick reference for releasing packages
Release Setup - Release infrastructure documentation

CLI Commands

Chat Command - Stateless AI interactions
Conversation Command - Stateful conversations with advanced features
Config Command - Configuration management and TypeScript benefits

Package Documentation

Core Packages:

@anygpt/types - Pure type definitions
@anygpt/config - Configuration management
@anygpt/router - Core router and connector system
@anygpt/ai-provider - AI provider wrapper with function calling
@anygpt/rules - Type-safe rule engine
@anygpt/mcp-logger - File-based logging for MCP servers
@anygpt/plugins - Plugin system for dynamic configuration

MCP & Discovery:

@anygpt/mcp-discovery - MCP tool discovery engine
@anygpt/mcp-discovery-server - MCP Discovery Server (PRIMARY interface)
@anygpt/mcp - MCP server implementation

Connectors:

@anygpt/openai - OpenAI connector
@anygpt/anthropic - Anthropic connector
@anygpt/claude - Claude via MCP
@anygpt/cody - Sourcegraph Cody
@anygpt/mock - Mock connector for testing

CLI:

@anygpt/cli - Command-line interface

Architecture Documentation

Router API Reference - Complete API documentation
Router Architecture - System design patterns
Configuration Guide - Provider configuration
Connector Usage - Provider-specific usage

CLI Documentation

CLI Overview - Complete CLI documentation
Tag Resolution Guide - How to use tags and model discovery
Chat Command - Stateless chat usage
Conversation Command - Stateful conversations
Config Command - Configuration management
Benchmark Command - Model performance testing

Security

⚠️ Important: This project handles sensitive credentials. Please review SECURITY.md before contributing.

Key security practices:

Never commit API keys or tokens
Use environment variables for credentials
Run security checks before committing (see .windsurf/workflows/security-check.md)
Use generic examples (e.g., example.com) instead of internal URLs

License

MIT License - see LICENSE file for details.