MCP Gateway
A Bun-based gateway that provides self-describing, discoverable APIs for AI agents, restoring the hypermedia pattern of the web. It allows agents to autonomously explore and understand API surfaces through MCP protocol.
README
MCP Gateway
A Bun-based gateway that brings back the original vision of the web: self-describing, discoverable APIs for AI agents.
Why This Exists
HTTP and HTML were designed as a hypermedia system—clients don't need external documentation because responses contain links describing available actions. REST APIs abandoned this, requiring separate docs for clients to know what endpoints exist.
MCP Gateway restores this pattern for AI agents:
| Original Web | MCP Gateway |
|---|---|
| Browser starts at URL | Agent connects to /mcp |
| HTML describes available links | Service cards describe available operations |
| User clicks links to navigate | Agent calls tools/list → tools/call |
| No external docs needed | No external docs needed |
The service card is the modern equivalent of an HTML page with links—a self-describing resource that tells the agent what it can do next.
How It Works
Agent Gateway
│ │
├── tools/list ────────────────►│
│◄─── [gateway, memory, mesh] ──│
│ │
├── tools/call memory ──────────►│
│◄─── "POST /semantic, GET /schema, ..." ──│
│ │
├── GET /memory/schema ─────────►│ (direct HTTP)
│◄─── { labels: [...] } ────────│
The agent autonomously explores and understands the API surface, just like a browser exploring a website.
Features
- Self-Describing APIs: Service discovery via MCP protocol (JSON-RPC 2.0)
- OpenAPI-Compatible Config: Define services using familiar OpenAPI structure
- Dynamic Service Loading: Drop TypeScript files in
services/and they're auto-loaded - STDIO Transport: Connect Claude Code directly via STDIO wrapper
- Zero Dependencies: Pure Bun, no external runtime dependencies
Note: The
services/directory contains example implementations (memory, mesh, recall, orchestrator) that demonstrate the gateway pattern. These are works in progress and require external backends (Neo4j, embedding service). The gateway itself is stable—use it with your own services.
Quick Start
# Clone and install
git clone https://github.com/HiddenDeveloper/mcp-gateway
cd mcp-gateway
bun install
# Start the gateway
bun run src/gateway.ts
Visit http://localhost:3000 to see the gateway landing page.
Usage
MCP Discovery
# List available services
curl -X POST http://localhost:3000/mcp \
-H "Content-Type: application/json" \
-d '{"jsonrpc": "2.0", "method": "tools/list", "id": 1}'
# Get service details
curl -X POST http://localhost:3000/mcp \
-H "Content-Type: application/json" \
-d '{"jsonrpc": "2.0", "method": "tools/call", "params": {"name": "memory_service_card"}, "id": 2}'
Direct HTTP Calls
# Call operations directly
curl http://localhost:3000/memory/schema
curl -X POST http://localhost:3000/memory/semantic -d '{"query": "search term"}'
Claude Code Integration
Add to your Claude Code MCP settings:
{
"mcpServers": {
"my-gateway": {
"command": "bun",
"args": ["run", "/path/to/mcp-gateway/src/stdio/wrapper.ts"]
}
}
}
The gateway must be running before using the wrapper.
Architecture
┌─────────────────────────────┐
│ Bun Gateway (:3000) │
│ │
MCP Client ────────►│ /mcp → MCP Handler │
(HTTP transport) │ (service discovery) │
│ │
Claude Code ───────►│ STDIO Wrapper ─────────────┼──► /mcp
(STDIO transport) │ (src/stdio/wrapper.ts) │
│ │
HTTP Client ───────►│ /{service}/{path} │
│ → Router │
│ → Service Loader │
│ → Service Handler │
└─────────────────────────────┘
│
▼
┌─────────────────────────────┐
│ services/ │
│ └── memory/ │
│ ├── get_schema.ts │
│ ├── semantic_search │
│ ├── text_search.ts │
│ └── execute_cypher │
└─────────────────────────────┘
Configuration
Config File
Create config/gateway.json:
{
"gateway": {
"info": { "title": "my-gateway", "version": "1.0.0" },
"service_card": {
"operationId": "gateway",
"summary": "Gateway service discovery"
}
},
"services": [{
"servers": [{ "url": "http://localhost:3000/memory" }],
"service_card": {
"operationId": "memory",
"summary": "Memory service"
},
"paths": {
"/schema": {
"get": {
"operationId": "get_schema",
"summary": "Get database schema"
}
}
}
}]
}
Environment Variables
| Variable | Default | Description |
|---|---|---|
MCP_CONFIG_PATH |
./config/gateway.json |
Config file path |
MCP_PORT |
3000 |
Server port |
MCP_SERVICES_DIR |
./services |
Services directory |
MCP_GATEWAY_URL |
http://localhost:3000/mcp |
Gateway URL (STDIO wrapper) |
MCP_DEBUG |
false |
Debug logging (STDIO wrapper) |
Writing Services
Services are TypeScript files that export a default async function:
// services/memory/semantic_search.ts
import { getNeo4jService } from "./lib/config";
interface Params {
query: string;
limit?: number;
threshold?: number;
}
export default async function(params: Params) {
const { query, limit = 10, threshold = 0.7 } = params;
const service = getNeo4jService();
const results = await service.semanticSearch(query, ["KnowledgeItem"], "embedding_vectors", limit);
return { matches: results.filter(r => r.score >= threshold) };
}
The gateway:
- Loads services from
services/{serviceId}/{operationId}.ts - Routes HTTP requests to the matching service
- Passes query params, path params, and body as
params - Returns the service result as JSON
MCP Protocol
| Method | Path | Description |
|---|---|---|
| GET | /mcp |
Server info and capabilities |
| POST | /mcp |
JSON-RPC 2.0 protocol |
| GET | /health |
Health check |
| GET | / |
Landing page |
Supported JSON-RPC Methods
initialize- Initialize MCP connectiontools/list- List available service cardstools/call- Get service details (returns operations)ping- Health check
Development
# Run tests
bun test
# Run specific tests
bun test tests/config.test.ts
bun test tests/mcp.test.ts
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。