内容
# tRPC MCP Go: Model Context Protocol Implementation with Streaming HTTP Support
A Go implementation of the [Model Context Protocol (MCP)](https://github.com/modelcontextprotocol/modelcontextprotocol) with comprehensive streaming HTTP support. This library enables efficient communication between client applications and tools/resources.
## Features
### Core Features
- **Full MCP Specification Support**: Implements MCP, supporting protocol versions up to 2025-03-26 (defaulting to 2024-11-05 for client compatibility in examples).
- **Streaming Support**: Real-time data streaming with Server-Sent Events (SSE)
- **Tool Framework**: Register and execute tools with structured parameter handling
- **Struct-First API**: Generate schemas automatically from Go structs with type safety
- **Resource Management**: Serve text and binary resources with RESTful interfaces
- **Prompt Templates**: Create and manage prompt templates for LLM interactions
- **Progress Notifications**: Built-in support for progress updates on long-running operations
- **Logging System**: Integrated logging with configurable levels
### Transport Options
trpc-mcp-go supports three transport mechanisms for different deployment scenarios:
#### **STDIO Transport**
- **Use Case**: Local integrations, command-line tools, cross-language compatibility
- **Communication**: Standard input/output streams with JSON-RPC
- **Process Model**: Client launches server as subprocess
- **Benefits**: Simple, cross-platform, supports TypeScript/Python servers
#### **Streamable HTTP Transport** (Recommended)
- **Use Case**: Web integrations, multi-client servers, production deployments
- **Communication**: HTTP POST requests with optional SSE streaming
- **Features**: Session management, resumable connections, scalability
- **Response Modes**:
- JSON responses for simple request-response
- SSE streaming for real-time operations
- **Server Notifications**: Optional GET SSE endpoint for server-initiated messages
#### **SSE Transport (Legacy)**
- **Use Case**: Backward compatibility with MCP protocol version 2024-11-05
- **Communication**: HTTP POST requests + dedicated SSE endpoint for server messages
- **Features**: Always stateful with session management and persistent SSE connections
- **Status**: Deprecated in favor of Streamable HTTP, maintained for compatibility
### Connection Modes
**Streamable HTTP Transport** supports multiple connection modes:
- **Stateless**: Simple request-response pattern without persistent sessions
- **Stateful**: Session management with user context and persistent connections
**SSE Transport** is always stateful by design with persistent connections.
**STDIO Transport** uses process-based communication (no HTTP sessions).
## Installation
```bash
go get trpc.group/trpc-go/trpc-mcp-go
```
## Quick Start
### Choose Your Transport
Pick the transport that best fits your use case:
| Transport | When to Use | Example |
|-----------|-------------|---------|
| **STDIO** | Local tools, CLI integration, cross-language compatibility | [`examples/transport-modes/stdio/`](examples/transport-modes/stdio/) |
| **Streamable HTTP** | Web services, multi-client servers, production apps | [`examples/quickstart/`](examples/quickstart/) |
| **STDIO to Streamable Proxy** | Expose an existing stdio MCP server as a remote Streamable HTTP MCP server | `NewStreamableServerWithStdio` |
| **SSE (Legacy)** | Compatibility with older MCP clients | [`examples/transport-modes/sse-legacy/`](examples/transport-modes/sse-legacy/) |
### Streamable HTTP Server Example (Recommended)
```go
package main
import (
"context"
"fmt"
"log"
"os"
"os/signal"
"syscall"
mcp "trpc.group/trpc-go/trpc-mcp-go"
)
func main() {
// Print startup message.
log.Printf("Starting basic example server...")
// Create server using the new API style:
// - First two required parameters: server name and version
// - WithServerAddress sets the address to listen on (default: "localhost:3000")
// - WithServerPath sets the API path prefix
// - WithServerLogger injects logger at the server level
mcpServer := mcp.NewServer(
"Basic-Example-Server",
"0.1.0",
mcp.WithServerAddress(":3000"),
mcp.WithServerPath("/mcp"),
mcp.WithServerLogger(mcp.GetDefaultLogger()),
)
// Register basic greet tool.
greetTool := mcp.NewTool("greet",
mcp.WithDescription("A simple greeting tool."),
mcp.WithString("name", mcp.Description("Name to greet.")))
greetHandler := func(ctx context.Context, req *mcp.CallToolRequest) (*mcp.CallToolResult, error) {
// Check if the context is cancelled.
select {
case <-ctx.Done():
return mcp.NewErrorResult("Request cancelled"), ctx.Err()
default:
// Continue execution.
}
// Extract name parameter.
name := "World"
if nameArg, ok := req.Params.Arguments["name"]; ok {
if nameStr, ok := nameArg.(string); ok && nameStr != "" {
name = nameStr
}
}
// Create greeting message.
greeting := fmt.Sprintf("Hello, %s!", name)
// Create tool result.
return mcp.NewTextResult(greeting), nil
}
mcpServer.RegisterTool(greetTool, greetHandler)
log.Printf("Registered basic greet tool: greet")
// Set up a graceful shutdown.
stop := make(chan os.Signal, 1)
signal.Notify(stop, os.Interrupt, syscall.SIGTERM)
// Start server (run in goroutine).
go func() {
log.Printf("MCP server started, listening on port 3000, path /mcp")
if err := mcpServer.Start(); err != nil {
log.Fatalf("Server failed to start: %v", err)
}
}()
// Wait for termination signal.
<-stop
log.Printf("Shutting down server...")
}
```
### Streamable HTTP Client Example
```go
package main
import (
"context"
"fmt"
"log"
"os"
mcp "trpc.group/trpc-go/trpc-mcp-go"
)
// initializeClient initializes the MCP client with server connection and session setup
func initializeClient(ctx context.Context) (*mcp.Client, error) {
log.Println("===== Initialize client =====")
serverURL := "http://localhost:3000/mcp"
mcpClient, err := mcp.NewClient(
serverURL,
mcp.Implementation{
Name: "MCP-Go-Client",
Version: "1.0.0",
},
mcp.WithClientLogger(mcp.GetDefaultLogger()),
)
if err != nil {
return nil, fmt.Errorf("failed to create client: %v", err)
}
initResp, err := mcpClient.Initialize(ctx, &mcp.InitializeRequest{})
if err != nil {
mcpClient.Close()
return nil, fmt.Errorf("initialization failed: %v", err)
}
log.Printf("Server info: %s %s", initResp.ServerInfo.Name, initResp.ServerInfo.Version)
log.Printf("Protocol version: %s", initResp.ProtocolVersion)
if initResp.Instructions != "" {
log.Printf("Server instructions: %s", initResp.Instructions)
}
sessionID := mcpClient.GetSessionID()
if sessionID != "" {
log.Printf("Session ID: %s", sessionID)
}
return mcpClient, nil
}
// handleTools manages tool-related operations including listing and calling tools
func handleTools(ctx context.Context, client *mcp.Client) error {
log.Println("===== List available tools =====")
listToolsResp, err := client.ListTools(ctx, &mcp.ListToolsRequest{})
if err != nil {
return fmt.Errorf("failed to list tools: %v", err)
}
tools := listToolsResp.Tools
if len(tools) == 0 {
log.Printf("No available tools.")
return nil
}
log.Printf("Found %d tools:", len(tools))
for _, tool := range tools {
log.Printf("- %s: %s", tool.Name, tool.Description)
}
// Call the first tool
log.Printf("===== Call tool: %s =====", tools[0].Name)
callToolReq := &mcp.CallToolRequest{}
callToolReq.Params.Name = tools[0].Name
callToolReq.Params.Arguments = map[string]interface{}{
"name": "MCP User",
}
callToolResp, err := client.CallTool(ctx, callToolReq)
if err != nil {
return fmt.Errorf("failed to call tool: %v", err)
}
log.Printf("Tool result:")
for _, item := range callToolResp.Content {
if textContent, ok := item.(mcp.TextContent); ok {
log.Printf(" %s", textContent.Text)
}
}
return nil
}
// terminateSession handles the termination of the current session
func terminateSession(ctx context.Context, client *mcp.Client) error {
sessionID := client.GetSessionID()
if sessionID == "" {
return nil
}
log.Printf("===== Terminate session =====")
if err := client.TerminateSession(ctx); err != nil {
return fmt.Errorf("failed to terminate session: %v", err)
}
log.Printf("Session terminated.")
return nil
}
func main() {
// Initialize log.
log.Println("Starting example client...")
// Create context.
ctx := context.Background()
// Initialize client
client, err := initializeClient(ctx)
if err != nil {
log.Printf("Error: %v\n", err)
os.Exit(1)
}
defer client.Close()
// Handle tools
if err := handleTools(ctx, client); err != nil {
log.Printf("Error: %v\n", err)
}
// Terminate session
if err := terminateSession(ctx, client); err != nil {
log.Printf("Error: %v\n", err)
}
log.Printf("Example finished.")
}
```
### STDIO Transport Example
For local integrations and cross-language compatibility:
**Server (server/main.go):**
```go
func main() {
// Create STDIO server
mcpServer := mcp.NewStdioServer("My-STDIO-Server", "1.0.0")
// Register tools
greetTool := mcp.NewTool("greet",
mcp.WithDescription("Greet someone"),
mcp.WithString("name", mcp.Description("Name to greet")))
mcpServer.RegisterTool(greetTool, func(ctx context.Context, req *mcp.CallToolRequest) (*mcp.CallToolResult, error) {
name := req.Params.Arguments["name"].(string)
return mcp.NewTextResult(fmt.Sprintf("Hello, %s!", name)), nil
})
// Start STDIO server (reads from stdin, writes to stdout)
if err := mcpServer.Start(); err != nil {
log.Fatalf("Server failed: %v", err)
}
}
```
**Client:**
```go
func main() {
// Create STDIO client that launches server as subprocess
client, err := mcp.NewStdioClient(
mcp.StdioTransportConfig{
ServerParams: mcp.StdioServerParameters{
Command: "go",
Args: []string{"run", "./server/main.go"},
},
},
mcp.Implementation{Name: "My-Client", Version: "1.0.0"},
)
if err != nil {
log.Fatal(err)
}
defer client.Close()
// Initialize and use the client
ctx := context.Background()
if _, err := client.Initialize(ctx, &mcp.InitializeRequest{}); err != nil {
log.Fatal(err)
}
// Call tools
result, err := client.CallTool(ctx, &mcp.CallToolRequest{
Params: mcp.CallToolParams{
Name: "greet",
Arguments: map[string]interface{}{"name": "World"},
},
})
if err != nil {
log.Fatal(err)
}
if result.IsError {
log.Fatalf("tool failed: %v", result.Content)
}
fmt.Printf("Result: %v\n", result.Content[0])
}
```
### STDIO to Streamable HTTP Proxy
Use `NewStreamableServerWithStdio` when you need to expose an existing stdio MCP server as a remote Streamable HTTP MCP server. The proxy starts the stdio process as an MCP client, initializes it, discovers its tools/resources/prompts, registers local forwarding handlers, and serves them through the normal Streamable HTTP `Server`.
```go
package main
import (
"context"
"log"
"os"
"os/signal"
"syscall"
"time"
mcp "trpc.group/trpc-go/trpc-mcp-go"
)
func main() {
ctx := context.Background()
server, proxy, err := mcp.NewStreamableServerWithStdio(ctx, mcp.StreamableStdioProxyConfig{
ServerName: "filesystem-proxy",
ServerVersion: "1.0.0",
Stdio: mcp.StdioTransportConfig{
ServerParams: mcp.StdioServerParameters{
Command: "npx",
Args: []string{"-y", "@modelcontextprotocol/server-filesystem", "/tmp"},
},
Timeout: 30 * time.Second,
},
ServerOptions: []mcp.ServerOption{
mcp.WithServerAddress(":3000"),
mcp.WithServerPath("/mcp"),
},
})
if err != nil {
log.Fatalf("failed to create stdio proxy: %v", err)
}
defer proxy.Close()
stop := make(chan os.Signal, 1)
signal.Notify(stop, os.Interrupt, syscall.SIGTERM)
go func() {
log.Printf("stdio MCP server is available as Streamable HTTP at http://localhost:3000/mcp")
if err := server.Start(); err != nil {
log.Fatalf("server failed: %v", err)
}
}()
<-stop
}
```
For simple cases, use the shorter helper:
```go
server, proxy, err := mcp.NewStreamableServerWithStdioParams(
ctx,
"stdio-proxy",
"1.0.0",
mcp.StdioTransportConfig{
ServerParams: mcp.StdioServerParameters{
Command: "go",
Args: []string{"run", "./server/main.go"},
},
},
mcp.WithServerAddress(":3000"),
mcp.WithServerPath("/mcp"),
)
```
This proxy follows the same wrapper pattern as the legacy stdio-to-SSE adapter: `tools/list`, `resources/list`, and `prompts/list` are discovered at startup, including paginated results, then `tools/call`, `resources/read`, and `prompts/get` are forwarded to the stdio process at call time. A proxy instance owns one stdio child process that is shared by all Streamable HTTP sessions. It is not a fully transparent JSON-RPC tunnel; unsupported custom methods and advanced server-to-client flows should be implemented explicitly.
## Configuration
### Server Configuration
The server can be configured using option functions:
```go
server := mcp.NewServer(
"My-MCP-Server", // Server name
"1.0.0", // Server version
mcp.WithServerAddress(":3000"), // Listen address
mcp.WithServerPath("/mcp"), // API path prefix
mcp.WithPostSSEEnabled(true), // Enable SSE responses
mcp.WithGetSSEEnabled(true), // Allow GET for SSE
mcp.WithStatelessMode(false), // Use stateful mode
mcp.WithServerLogger(mcp.GetDefaultLogger()), // Custom logger
)
```
### Available Server Options
| Option | Description | Default |
|--------|-------------|---------|
| `WithServerAddress` | Set server address to listen on | `"localhost:3000"` |
| `WithServerPath` | Set API path prefix | `/mcp` |
| `WithServerLogger` | Custom logger for server | Default logger |
| `WithoutSession` | Disable session management | Sessions enabled |
| `WithPostSSEEnabled` | Enable SSE responses | `true` |
| `WithGetSSEEnabled` | Allow GET for SSE connections | `true` |
| `WithNotificationBufferSize` | Size of notification buffer | `10` |
| `WithStatelessMode` | Run in stateless mode | `false` |
### Client Configuration
The client can be configured using option functions:
```go
client, err := mcp.NewClient(
"http://localhost:3000/mcp", // Server URL
mcp.Implementation{ // Client info
Name: "MCP-Client",
Version: "1.0.0",
},
mcp.WithProtocolVersion(mcp.ProtocolVersion_2025_03_26), // Protocol version
mcp.WithClientGetSSEEnabled(true), // Use GET for SSE
mcp.WithClientLogger(mcp.GetDefaultLogger()), // Custom logger
)
```
### Available Client Options
| Option | Description | Default |
|--------|-------------|---------|
| `WithProtocolVersion` | Specify MCP protocol version | `mcp.ProtocolVersion_2025_03_26` |
| `WithClientGetSSEEnabled` | Use GET for SSE instead of POST | `true` |
| `WithClientLogger` | Custom logger for client | Default logger |
| `WithClientPath` | Set custom client path | Server path |
| `WithHTTPReqHandler` | Use custom HTTP request handler | Default handler |
## Advanced Features
### Streaming Progress with SSE
Create tools that provide real-time progress updates:
```go
// handleMultiStageGreeting handles the multi-stage greeting tool and sends multiple notifications via SSE.
func handleMultiStageGreeting(ctx context.Context, req *mcp.CallToolRequest) (*mcp.CallToolResult, error) {
// Extract name from parameters.
name := "Guest"
if nameArg, ok := req.Params.Arguments["name"]; ok {
if nameStr, ok := nameArg.(string); ok && nameStr != "" {
name = nameStr
}
}
stages := 3
if stagesArg, ok := req.Params.Arguments["stages"]; ok {
if stagesFloat, ok := stagesArg.(float64); ok && stagesFloat > 0 {
stages = int(stagesFloat)
}
}
// Get notification sender from context.
notificationSender, hasNotificationSender := mcp.GetNotificationSender(ctx)
if !hasNotificationSender {
log.Printf("unable to get notification sender from context")
return &mcp.CallToolResult{
Content: []mcp.Content{
mcp.NewTextContent("Error: unable to get notification sender."),
},
}, fmt.Errorf("unable to get notification sender from context")
}
// Send progress update.
sendProgress := func(progress float64, message string) {
err := notificationSender.SendProgress(progress, message)
if err != nil {
log.Printf("Failed to send progress notification: %v", err)
}
}
// Send log message.
sendLogMessage := func(level string, message string) {
err := notificationSender.SendLogMessage(level, message)
if err != nil {
log.Printf("Failed to send log notification: %v", err)
}
}
// Start greeting process.
sendProgress(0.0, "Start multi-stage greeting")
sendLogMessage("info", fmt.Sprintf("Start greeting to %s", name))
time.Sleep(500 * time.Millisecond)
// Send multiple stage notifications.
for i := 1; i <= stages; i++ {
// Check if context is canceled.
select {
case <-ctx.Done():
return &mcp.CallToolResult{
Content: []mcp.Content{
mcp.NewTextContent(fmt.Sprintf("Greeting canceled at stage %d", i)),
},
}, ctx.Err()
default:
// Continue sending.
}
sendProgress(float64(i)/float64(stages), fmt.Sprintf("Stage %d greeting", i))
sendLogMessage("info", fmt.Sprintf("Stage %d greeting: Hello %s!", i, name))
time.Sleep(800 * time.Millisecond)
}
// Send final greeting.
sendProgress(1.0, "Greeting completed")
sendLogMessage("info", fmt.Sprintf("Completed multi-stage greeting to %s", name))
// Return final result.
return &mcp.CallToolResult{
Content: []mcp.Content{
mcp.NewTextContent(fmt.Sprintf(
"Completed %d-stage greeting to %s!",
stages, name,
)),
},
}, nil
}
func main() {
// Create MCP server
mcpServer := mcp.NewServer(
"Multi-Stage-Greeting-Server",
"1.0.0",
mcp.WithServerAddress(":3000"),
mcp.WithServerPath("/mcp"),
mcp.WithPostSSEEnabled(true),
)
// Register a multi-stage greeting tool
multiStageGreetingTool := mcp.NewTool("multi-stage-greeting",
mcp.WithDescription("Send multi-stage greeting via SSE."),
mcp.WithString("name", mcp.Description("Name to greet.")),
mcp.WithNumber("stages",
mcp.Description("Number of greeting stages."),
mcp.Default(3),
),
)
mcpServer.RegisterTool(multiStageGreetingTool, handleMultiStageGreeting)
log.Printf("Registered multi-stage greeting tool: multi-stage-greeting")
// Start server
log.Printf("MCP server started at :3000, path /mcp")
if err := mcpServer.Start(); err != nil {
log.Fatalf("Failed to start server: %v", err)
}
}
```
### Client-Side Progress Handling
```go
package main
import (
"context"
"fmt"
"log"
"trpc.group/trpc-go/trpc-mcp-go"
)
// Example NotificationCollector structure and methods
type NotificationCollector struct{}
func (nc *NotificationCollector) HandleProgress(notification *mcp.JSONRPCNotification) error {
progress, _ := notification.Params.AdditionalFields["progress"].(float64)
message, _ := notification.Params.AdditionalFields["message"].(string)
fmt.Printf("Progress: %.0f%% - %s\n", progress*100, message)
return nil
}
func (nc *NotificationCollector) HandleLog(notification *mcp.JSONRPCNotification) error {
level, _ := notification.Params.AdditionalFields["level"].(string)
data, _ := notification.Params.AdditionalFields["data"].(string)
fmt.Printf("[%s] %s\n", level, data)
return nil
}
func main() {
// Create context
ctx := context.Background()
// Initialize client
client, err := mcp.NewClient(
"http://localhost:3000/mcp",
mcp.Implementation{
Name: "MCP-Client-Stream-Handler",
Version: "1.0.0",
},
mcp.WithClientGetSSEEnabled(true),
)
if err != nil {
log.Fatalf("Failed to create client: %v", err)
}
defer client.Close()
// Initialize connection
_, err = client.Initialize(ctx, &mcp.InitializeRequest{})
if err != nil {
log.Fatalf("Failed to initialize client: %v", err)
}
// Create notification collector
collector := &NotificationCollector{}
// Register notification handlers
client.RegisterNotificationHandler("notifications/progress", collector.HandleProgress)
client.RegisterNotificationHandler("notifications/message", collector.HandleLog)
// Call tool with streaming
log.Printf("Calling multi-stage greeting tool...")
callRes, err := client.CallTool(ctx, &mcp.CallToolRequest{
Params: mcp.CallToolParams{
Name: "multi-stage-greeting",
Arguments: map[string]interface{}{
"name": "MCP User",
"stages": 5,
},
},
})
if err != nil {
log.Printf("Tool call failed: %v", err)
return
}
// Process final result
for _, item := range callRes.Content {
if textContent, ok := item.(mcp.TextContent); ok {
log.Printf("Final tool result: %s", textContent.Text)
}
}
log.Printf("Client example finished.")
}
```
### Resource Management
#### Single Content Resources
Register and serve resources with single content:
```go
// Register text resource
textResource := &mcp.Resource{
URI: "resource://example/text",
Name: "example-text",
Description: "Example text resource",
MimeType: "text/plain",
}
// Define text resource handler
textHandler := func(ctx context.Context, req *mcp.ReadResourceRequest) (mcp.ResourceContents, error) {
return mcp.TextResourceContents{
URI: textResource.URI,
MIMEType: textResource.MimeType,
Text: "This is an example text resource content.",
}, nil
}
// Register the text resource
server.RegisterResource(textResource, textHandler)
```
#### Multiple Contents Resources (Recommended)
Register resources that can provide multiple content representations:
```go
// Register a resource with multiple content formats
multiResource := &mcp.Resource{
URI: "resource://example/document",
Name: "example-document",
Description: "Document available in multiple formats",
MimeType: "text/plain",
}
// Define multiple contents handler
multiHandler := func(ctx context.Context, req *mcp.ReadResourceRequest) ([]mcp.ResourceContents, error) {
return []mcp.ResourceContents{
// Text representation
mcp.TextResourceContents{
URI: req.Params.URI,
MIMEType: "text/plain",
Text: "Document content in plain text format.",
},
// JSON representation
mcp.BlobResourceContents{
URI: req.Params.URI,
MIMEType: "application/json",
Blob: "eyJjb250ZW50IjogIkRvY3VtZW50IGNvbnRlbnQgaW4gSlNPTiBmb3JtYXQuIn0=", // base64 encoded JSON
},
}, nil
}
// Register the resource with multiple contents (recommended)
server.RegisterResources(multiResource, multiHandler)
```
The `RegisterResources` method is recommended as it:
- Aligns with the MCP protocol specification
- Allows a single resource to provide multiple content representations
- Supports use cases like serving the same data in different formats (text, JSON, XML, etc.)
### Prompt
Register prompt:
```go
// Register basic prompt
basicPrompt := &mcp.Prompt{
Name: "basic-prompt",
Description: "Basic prompt example",
Arguments: []mcp.PromptArgument{
{
Name: "name",
Description: "User name",
Required: true,
},
},
}
// Define basic prompt handler
basicPromptHandler := func(ctx context.Context, req *mcp.GetPromptRequest) (*mcp.GetPromptResult, error) {
name := req.Params.Arguments["name"]
return &mcp.GetPromptResult{
Description: basicPrompt.Description,
Messages: []mcp.PromptMessage{
{
Role: "user",
Content: mcp.TextContent{
Type: "text",
Text: fmt.Sprintf("Hello, %s! This is a basic prompt example.", name),
},
},
},
}, nil
}
// Register the basic prompt
server.RegisterPrompt(basicPrompt, basicPromptHandler)
```
## Struct-First API (Recommended)
Define MCP tools using Go structs for automatic schema generation and type safety:
```go
// Define input/output structures
type WeatherInput struct {
Location string `json:"location" jsonschema:"required,description=City name"`
Units string `json:"units,omitempty" jsonschema:"description=Temperature units,enum=celsius,enum=fahrenheit,default=celsius"`
}
type WeatherOutput struct {
Temperature float64 `json:"temperature" jsonschema:"description=Current temperature"`
Description string `json:"description" jsonschema:"description=Weather description"`
}
// Create tool with automatic schema generation
weatherTool := mcp.NewTool(
"get_weather",
mcp.WithDescription("Get weather information"),
mcp.WithInputStruct[WeatherInput](), // 🚀 Auto-generate input schema
mcp.WithOutputStruct[WeatherOutput](), // 🚀 Auto-generate output schema
)
// Type-safe handler with automatic validation
weatherHandler := mcp.NewTypedToolHandler(func(ctx context.Context, req *mcp.CallToolRequest, input WeatherInput) (WeatherOutput, error) {
return WeatherOutput{
Temperature: 22.5,
Description: "Partly cloudy",
}, nil
})
```
**Schema Generation Styles:**
By default, schemas use `$defs + $ref` for compact representation (~2KB). You can customize the generation style:
```go
// Ref style (default) - compact schemas with $defs + $ref
mcp.WithInputStruct[Input](mcp.WithRefStyle())
// Inline style - expand all types inline (larger schemas)
mcp.WithOutputStruct[Output](mcp.WithInlineStyle())
```
**Custom Schemas:**
When a schema comes from configuration or another schema source, pass the
complete OpenAPI schema directly instead of rebuilding it property by property:
```go
var inputSchema openapi3.Schema
if err := json.Unmarshal(schemaJSON, &inputSchema); err != nil {
return err
}
configuredTool := mcp.NewTool(
"configured_tool",
mcp.WithDescription("Tool whose input schema comes from configuration"),
mcp.WithInputSchema(&inputSchema),
)
```
`WithInputSchema` replaces the default input schema, while `WithOutputSchema`
sets the structured output schema. These options change the schema advertised
to MCP clients; the handler must still parse, validate, and process compatible
arguments and results.
The root input schema must have `type: "object"`, as required by MCP. Tool
options are applied in order: property options such as `WithString` can extend
a custom input schema, while a later `WithInputSchema` replaces earlier input
schema changes. The tool retains the provided schema pointer, so treat it as
immutable after `NewTool` returns; sharing the pointer is safe only while it
remains immutable.
**Benefits:**
- 🛡️ **Type Safety**: Compile-time type checking and automatic validation
- 📋 **Rich Schemas**: Auto-generated OpenAPI schemas with descriptions, enums, defaults
- 🔄 **Structured Output**: Type-safe responses with backward compatibility
- 🎯 **DRY Principle**: Single source of truth for data structures
- 🎨 **Flexible Styles**: Choose between compact ($ref) or expanded (inline) schemas
See [`examples/schema-generation/`](examples/schema-generation/) for a complete example.
## Example Patterns
The project includes several example patterns organized by category:
### Getting Started
| Pattern | Description |
|---------|-------------|
| [`quickstart/`](examples/quickstart/) | **Quick Start** - Simple tool registration and basic MCP usage |
### Transport Modes
| Pattern | Description |
|---------|-------------|
| [`transport-modes/stdio/`](examples/transport-modes/stdio/) | **STDIO Transport** - Cross-language compatibility with stdio |
| [`stdio-to-streamable-proxy/`](examples/stdio-to-streamable-proxy/) | **STDIO to Streamable Proxy** - Expose an existing stdio MCP server as Streamable HTTP |
| [`transport-modes/sse-legacy/`](examples/transport-modes/sse-legacy/) | **SSE Transport** - Server-Sent Events (legacy, for compatibility) |
| [`transport-modes/streamable-http/`](examples/transport-modes/streamable-http/) | **Streamable HTTP** - Various configuration options |
### ⚙Streamable HTTP Configurations
| Pattern | Response Mode | GET SSE | Session | Description |
|---------|---------------|---------|---------|-------------|
| [`stateless-json/`](examples/transport-modes/streamable-http/stateless-json/) | JSON Response | ❌ | ❌ | Simple request-response |
| [`stateful-json/`](examples/transport-modes/streamable-http/stateful-json/) | JSON Response | ❌ | ✅ | JSON with session management |
| [`stateless-sse/`](examples/transport-modes/streamable-http/stateless-sse/) | SSE Streaming | ❌ | ❌ | Real-time streaming responses |
| [`stateful-sse/`](examples/transport-modes/streamable-http/stateful-sse/) | SSE Streaming | ❌ | ✅ | Streaming with session management |
| [`stateful-json-getsse/`](examples/transport-modes/streamable-http/stateful-json-getsse/) | JSON Response | ✅ | ✅ | JSON + server notifications |
| [`stateful-sse-getsse/`](examples/transport-modes/streamable-http/stateful-sse-getsse/) | SSE Streaming | ✅ | ✅ | Full-featured streaming |
### Advanced Features
| Pattern | Description |
|---------|-------------|
| [`schema-generation/`](examples/schema-generation/) | **Struct-first API** - Auto-generate schemas from Go structs |
| [`resources-and-prompts/`](examples/resources-and-prompts/) | **Resources & Prompts** - Resource management and prompt templates |
| [`roots-management/`](examples/roots-management/) | **Roots Management** - Client filesystem integration across transports |
## FAQ
### 1. How to handle HTTP Headers?
**Q: How can I extract HTTP headers on the server side and send custom headers from the client?**
**A:** The library provides comprehensive HTTP header support for both server and client sides while maintaining transport layer independence.
#### Server Side: Extracting HTTP Headers
Use `WithHTTPContextFunc` to extract HTTP headers and make them available in tool handlers through context:
```go
package main
import (
"context"
"fmt"
"log"
"net/http"
mcp "trpc.group/trpc-go/trpc-mcp-go"
)
// Define context keys for type safety
type contextKey string
const (
AuthTokenKey contextKey = "auth_token"
UserAgentKey contextKey = "user_agent"
)
// HTTP context functions to extract headers
func extractAuthToken(ctx context.Context, r *http.Request) context.Context {
if authHeader := r.Header.Get("Authorization"); authHeader != "" {
return context.WithValue(ctx, AuthTokenKey, authHeader)
}
return ctx
}
func extractUserAgent(ctx context.Context, r *http.Request) context.Context {
if userAgent := r.Header.Get("User-Agent"); userAgent != "" {
return context.WithValue(ctx, UserAgentKey, userAgent)
}
return ctx
}
// Tool handler that accesses headers via context
func myTool(ctx context.Context, req *mcp.CallToolRequest) (*mcp.CallToolResult, error) {
// Extract headers from context
authToken, _ := ctx.Value(AuthTokenKey).(string)
userAgent, _ := ctx.Value(UserAgentKey).(string)
// Use header information in your tool logic
response := fmt.Sprintf("Received auth: %s, user-agent: %s", authToken, userAgent)
return mcp.NewTextResult(response), nil
}
func main() {
// Create server with HTTP context functions
server := mcp.NewServer(
"header-server", "1.0.0",
// Register multiple HTTP context functions
mcp.WithHTTPContextFunc(extractAuthToken),
mcp.WithHTTPContextFunc(extractUserAgent),
)
// Register tool
tool := mcp.NewTool("my-tool", mcp.WithDescription("Tool that uses headers"))
server.RegisterTool(tool, myTool)
// Start server
server.Start()
}
```
#### Client Side: Sending Custom HTTP Headers
**Option 1: Static Headers with `WithHTTPHeaders`**
Use `WithHTTPHeaders` for headers that don't change across requests:
```go
// Create custom headers
headers := make(http.Header)
headers.Set("Authorization", "Bearer your-token-here")
headers.Set("User-Agent", "MyMCPClient/1.0")
// Create client with static headers
client, err := mcp.NewClient(
"http://localhost:3000/mcp",
mcp.Implementation{Name: "my-client", Version: "1.0.0"},
mcp.WithHTTPHeaders(headers), // Applied to all requests
)
```
**Option 2: Dynamic Headers with `WithHTTPBeforeRequest`**
Use `WithHTTPBeforeRequest` for headers that change per request (e.g., per-request auth tokens, request IDs, tracing):
```go
// Define context keys
type contextKey string
const (
requestIDKey contextKey = "request-id"
userIDKey contextKey = "user-id"
)
// Create before-request function
beforeRequest := func(ctx context.Context, req *http.Request) error {
// Extract values from context and set headers
if requestID, ok := ctx.Value(requestIDKey).(string); ok {
req.Header.Set("X-Request-ID", requestID)
}
if userID, ok := ctx.Value(userIDKey).(string); ok {
req.Header.Set("X-User-ID", userID)
}
return nil // Return error to abort the request
}
// Create client with dynamic headers
client, err := mcp.NewClient(
"http://localhost:3000/mcp",
mcp.Implementation{Name: "my-client", Version: "1.0.0"},
mcp.WithHTTPBeforeRequest(beforeRequest),
)
// Pass context with values for each request
ctx := context.WithValue(context.Background(), requestIDKey, "req-12345")
ctx = context.WithValue(ctx, userIDKey, "user-789")
// Headers are dynamically set from context
result, err := client.CallTool(ctx, &mcp.CallToolRequest{
Params: mcp.CallToolParams{
Name: "my-tool",
Arguments: map[string]interface{}{"message": "Hello!"},
},
})
```
**Key Differences:**
| Feature | `WithHTTPHeaders` | `WithHTTPBeforeRequest` |
|---------|-------------------|-------------------------|
| **Use Case** | Static headers (API keys, service name) | Dynamic headers (request ID, user context) |
| **When Set** | Once at client creation | Before each HTTP request |
| **Context Access** | ❌ No | ✅ Yes - read from `context.Context` |
| **Applies To** | All requests | All requests (initialize, tools/list, tools/call, GET SSE) |
| **Can Abort Request** | ❌ No | ✅ Yes - return error |
**Combining Both:**
```go
// Static headers for service identification
headers := make(http.Header)
headers.Set("X-Service-Name", "my-service")
// Dynamic headers for per-request data
beforeRequest := func(ctx context.Context, req *http.Request) error {
if requestID, ok := ctx.Value(requestIDKey).(string); ok {
req.Header.Set("X-Request-ID", requestID)
}
return nil
}
client, err := mcp.NewClient(
serverURL,
clientInfo,
mcp.WithHTTPHeaders(headers), // Static headers
mcp.WithHTTPBeforeRequest(beforeRequest), // Dynamic headers
)
// Both are applied: static headers first, then dynamic headers
```
#### Key Features
- **Transport Layer Independence**: Tool handlers access headers through context, not directly from HTTP requests
- **Multiple Context Functions**: Use multiple `WithHTTPContextFunc` calls to extract different headers
- **Type Safety**: Use strongly-typed context keys to avoid conflicts
- **Static + Dynamic Headers**: Combine `WithHTTPHeaders` and `WithHTTPBeforeRequest` for maximum flexibility
- **Context Propagation**: `WithHTTPBeforeRequest` enables per-request headers via `context.Context`
- **All Requests Covered**: Headers apply to initialize, tools/list, tools/call, and GET SSE connections
- **Backward Compatibility**: Optional features that don't break existing APIs
#### Complete Example
See [`examples/quickstart/`](examples/quickstart/) for basic usage, or [`examples/transport-modes/`](examples/transport-modes/) for transport-specific header handling.
### 2. How to get server name and version in a tool handler?
You can retrieve the `mcp.Server` instance from the `context.Context` within your tool handler and then call its `GetServerInfo()` method:
```go
func handleMyTool(ctx context.Context, req *mcp.CallToolRequest) (*mcp.CallToolResult, error) {
serverInstance := mcp.GetServerFromContext(ctx)
if mcpServer, ok := serverInstance.(*mcp.Server); ok {
serverInfo := mcpServer.GetServerInfo()
// Now you have serverInfo.Name and serverInfo.Version
return mcp.NewTextResult(fmt.Sprintf("Server Name: %s, Version: %s", serverInfo.Name, serverInfo.Version)), nil
}
return mcp.NewTextResult("Could not retrieve server information."), nil
}
```
### 3. How to dynamically filter tools, prompts, and resources based on user context?
The library provides a powerful filtering system that works with HTTP header extraction to enable dynamic access control. You can filter tools, prompts, and resources based on user context such as roles, permissions, client version, or any other criteria.
#### Basic Context-Based Filtering
Use filtering options combined with `WithHTTPContextFunc` to filter tools, prompts, and resources based on user context (e.g., roles, permissions, version, etc.):
```go
package main
import (
"context"
"net/http"
mcp "trpc.group/trpc-go/trpc-mcp-go"
)
// Extract user role from HTTP headers
func extractUserRole(ctx context.Context, r *http.Request) context.Context {
if userRole := r.Header.Get("X-User-Role"); userRole != "" {
ctx = context.WithValue(ctx, "user_role", userRole)
}
return ctx
}
// Create role-based filter for tools
func createRoleBasedFilter() mcp.ToolListFilter {
return func(ctx context.Context, tools []*mcp.Tool) []*mcp.Tool {
userRole := "guest" // default
if role, ok := ctx.Value("user_role").(string); ok && role != "" {
userRole = role
}
var filtered []*mcp.Tool
for _, tool := range tools {
switch userRole {
case "admin":
filtered = append(filtered, tool) // Admin sees all tools
case "user":
if tool.Name == "calculator" || tool.Name == "weather" {
filtered = append(filtered, tool)
}
case "guest":
if tool.Name == "calculator" {
filtered = append(filtered, tool)
}
}
}
return filtered
}
}
// Create role-based filter for prompts
func createPromptFilter() mcp.PromptListFilter {
return func(ctx context.Context, prompts []*mcp.Prompt) []*mcp.Prompt {
userRole := "guest" // default
if role, ok := ctx.Value("user_role").(string); ok && role != "" {
userRole = role
}
var filtered []*mcp.Prompt
for _, prompt := range prompts {
switch userRole {
case "admin":
filtered = append(filtered, prompt) // Admin sees all prompts
case "user":
if prompt.Name == "greeting" || prompt.Name == "summary" {
filtered = append(filtered, prompt)
}
case "guest":
if prompt.Name == "greeting" {
filtered = append(filtered, prompt)
}
}
}
return filtered
}
}
// Create role-based filter for resources
func createResourceFilter() mcp.ResourceListFilter {
return func(ctx context.Context, resources []*mcp.Resource) []*mcp.Resource {
userRole := "guest" // default
if role, ok := ctx.Value("user_role").(string); ok && role != "" {
userRole = role
}
var filtered []*mcp.Resource
for _, resource := range resources {
switch userRole {
case "admin":
filtered = append(filtered, resource) // Admin sees all resources
case "user":
if resource.Name == "public_data" || resource.Name == "user_data" {
filtered = append(filtered, resource)
}
case "guest":
if resource.Name == "public_data" {
filtered = append(filtered, resource)
}
}
}
return filtered
}
}
func main() {
// Create server with context-based filtering for all types
server := mcp.NewServer(
"filtered-server", "1.0.0",
mcp.WithHTTPContextFunc(extractUserRole),
mcp.WithToolListFilter(createRoleBasedFilter()),
mcp.WithPromptListFilter(createPromptFilter()),
mcp.WithResourceListFilter(createResourceFilter()),
)
// Register tools, prompts, and resources as usual
server.RegisterTool(calculatorTool, calculatorHandler)
server.RegisterTool(weatherTool, weatherHandler)
server.RegisterTool(adminTool, adminHandler)
server.RegisterPrompt(userPrompt, userPromptHandler)
server.RegisterPrompt(adminPrompt, adminPromptHandler)
server.RegisterResource(publicResource, publicResourceHandler)
server.RegisterResource(adminResource, adminResourceHandler)
server.Start()
}
```
#### Advanced Custom Filter Functions
You can create more complex filtering logic:
```go
func createAdvancedFilter() mcp.ToolListFilter {
return func(ctx context.Context, tools []*mcp.Tool) []*mcp.Tool {
// Extract user information from context
userRole := "guest"
if role, ok := ctx.Value("user_role").(string); ok && role != "" {
userRole = role
}
clientVersion := ""
if version, ok := ctx.Value("client_version").(string); ok {
clientVersion = version
}
// Complex business logic
if userRole == "premium" && clientVersion == "2.0.0" {
return tools // Premium users with v2.0.0 see all tools
}
var filtered []*mcp.Tool
for _, tool := range tools {
if shouldShowTool(tool, userRole, clientVersion) {
filtered = append(filtered, tool)
}
}
return filtered
}
}
// Helper function for business logic
func shouldShowTool(tool *mcp.Tool, userRole, clientVersion string) bool {
// Implement your business logic here
switch tool.Name {
case "calculator":
return true // Available to everyone
case "weather":
return userRole == "user" || userRole == "admin"
case "admin_panel":
return userRole == "admin"
default:
return false
}
}
```
#### SSE Server Filtering
For SSE servers, use the corresponding SSE filter options:
```go
// Create SSE server with filtering
sseServer := mcp.NewSSEServer(
"filtered-sse-server", "1.0.0",
mcp.WithSSEContextFunc(extractUserRole),
mcp.WithSSEToolListFilter(createRoleBasedFilter()),
mcp.WithSSEPromptListFilter(createPromptFilter()),
mcp.WithSSEResourceListFilter(createResourceFilter()),
)
```
#### Testing Filtering
You can test different user contexts by sending HTTP headers:
```bash
# Admin sees all tools
curl -X POST http://localhost:3000/mcp \
-H "Content-Type: application/json" \
-H "X-User-Role: admin" \
-d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}'
# Test prompts filtering
curl -X POST http://localhost:3000/mcp \
-H "Content-Type: application/json" \
-H "X-User-Role: user" \
-d '{"jsonrpc": "2.0", "id": 2, "method": "prompts/list"}'
# Test resources filtering
curl -X POST http://localhost:3000/mcp \
-H "Content-Type: application/json" \
-H "X-User-Role: guest" \
-d '{"jsonrpc": "2.0", "id": 3, "method": "resources/list"}'
# User sees filtered tools
curl -X POST http://localhost:3000/mcp \
-H "Content-Type: application/json" \
-H "X-User-Role: user" \
-d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}'
```
#### Key Benefits
- **Flexible Filtering**: Create custom filters based on any context information
- **Transport Independence**: Works with both stateful and stateless modes
- **Performance**: Filters run only when tools are listed, not on every request
- **Security**: Tools are completely hidden from unauthorized users
### 4. How to configure retry mechanism for network errors?
The library provides automatic retry functionality at the transport layer to handle temporary network failures.
#### Simple Retry Configuration
```go
// Enable retry with 3 attempts (recommended for most use cases)
client, err := mcp.NewClient(serverURL, clientInfo,
mcp.WithSimpleRetry(3),
)
```
#### Advanced Retry Configuration
```go
// Custom retry configuration for specific scenarios
client, err := mcp.NewClient(serverURL, clientInfo,
mcp.WithRetry(mcp.RetryConfig{
MaxRetries: 5, // Maximum retry attempts
InitialBackoff: 1 * time.Second, // Initial delay before first retry
BackoffFactor: 1.5, // Exponential backoff multiplier
MaxBackoff: 15 * time.Second, // Maximum delay cap
}),
)
```
#### What Errors Are Retried?
The retry mechanism automatically handles:
- **Connection errors**: `connection refused`, `connection reset`, `connection timeout`
- **I/O timeouts**: `i/o timeout`, `read timeout`, `dial timeout`
- **Network errors**: `EOF`, `broken pipe`
- **HTTP server errors**: Status codes 408, 409, 429, and all 5xx errors
#### Key Features
- **Transport Layer**: Retry works across all MCP operations (tools, resources, prompts)
- **Exponential Backoff**: Intelligent delay strategy to avoid overwhelming servers
- **Error Classification**: Only retries temporary failures, not permanent errors (auth, bad requests)
- **Silent Operation**: No logging noise by default, retry happens transparently
- **Network Transports**: Works with Streamable HTTP and SSE transports (STDIO doesn't use retry due to its process-based nature)
### 5. How to use tool annotations for better UX?
Tool annotations provide behavioral hints to MCP clients about tool characteristics, enabling smarter usage and better user experience:
```go
tool := mcp.NewTool("weather_api",
mcp.WithDescription("Get weather information"),
mcp.WithString("location", mcp.Description("City name"), mcp.Required()),
mcp.WithToolAnnotations(&mcp.ToolAnnotations{
Title: "Weather Information", // Human-readable title
ReadOnlyHint: mcp.BoolPtr(true), // Safe, no side effects
DestructiveHint: mcp.BoolPtr(false), // Non-destructive operation
IdempotentHint: mcp.BoolPtr(true), // Same input → same output
OpenWorldHint: mcp.BoolPtr(true), // Interacts with external APIs
}),
)
```
**Key annotation types:**
- **ReadOnlyHint**: Tool doesn't modify environment (safe for repeated calls)
- **DestructiveHint**: Tool may cause permanent changes (only meaningful when ReadOnlyHint is false)
- **IdempotentHint**: Multiple calls with same arguments have same effect
- **OpenWorldHint**: Tool interacts with external systems vs internal data
> **Note**: Annotations are hints for UX optimization, not security guarantees. See `examples/annotations/` for comprehensive usage patterns.
### 6. How to get registered tools from server?
**Q: How can I check what tools are currently registered on the server before performing operations?**
**A:** Use the `GetTool` and `GetTools` methods available on all server types:
```go
// Check if a specific tool exists
if tool, exists := server.GetTool("weather_api"); exists {
fmt.Printf("Tool found: %s - %s\n", tool.Name, tool.Description)
// Safe to proceed with tool-related operations
}
// Get all registered tools
tools := server.GetTools()
fmt.Printf("Server has %d tools registered:\n", len(tools))
for _, tool := range tools {
fmt.Printf("- %s: %s\n", tool.Name, tool.Description)
}
```
**Key features:**
- **Safe Access**: Returns tool copies to prevent accidental modification
- **Universal Support**: Available on `Server`, `SSEServer`, and `StdioServer`
- **Dynamic Management**: Perfect for conditional tool registration/removal
- **Go Idioms**: Uses `(value, bool)` pattern for existence checking
### 7. How to use Middleware to intercept and modify requests/responses?
**Q: How can I add custom logic before/after tool execution, intercept specific requests, or modify results?**
**A:** The library provides a powerful middleware system that allows you to intercept and modify any MCP request/response. Middleware is especially useful for:
- **Tool/Prompt interception**: Mock, cache, or block specific tools/prompts
- **Result modification**: Transform or enhance responses
- **Custom result construction**: Return completely custom results without calling the actual handler
- **Cross-cutting concerns**: Logging, metrics, authentication, tracing
#### Basic Middleware Pattern
Middleware follows a simple function signature:
```go
type Middleware func(next HandlerFunc) HandlerFunc
type HandlerFunc func(ctx context.Context, req *JSONRPCRequest) (JSONRPCMessage, error)
```
Register middleware using `WithMiddleware` option:
```go
server := mcp.NewServer(
"my-server", "1.0.0",
mcp.WithMiddleware(
LoggingMiddleware,
MetricsMiddleware,
AuthMiddleware,
ToolInterceptorMiddleware,
),
)
```
Middleware executes in **onion model** (first registered = outer layer):
```
Request → Logging → Metrics → Auth → Tool Interceptor → Handler
↓
Response ← Logging ← Metrics ← Auth ← Tool Interceptor ← Handler
```
#### Example 1: Intercepting and Mocking Tool Calls
```go
// ToolInterceptorMiddleware intercepts specific tools and returns mock results
func ToolInterceptorMiddleware(next mcp.HandlerFunc) mcp.HandlerFunc {
return func(ctx context.Context, req *mcp.JSONRPCRequest) (mcp.JSONRPCMessage, error) {
// Only intercept tool call requests
if req.Method != mcp.MethodToolsCall {
return next(ctx, req)
}
// Parse tool name from request
var callReq mcp.CallToolRequest
if params, ok := req.Params.(map[string]interface{}); ok {
if name, ok := params["name"].(string); ok {
callReq.Params.Name = name
}
}
// Intercept specific tool and return mock result
if callReq.Params.Name == "expensive-api" {
log.Printf("🔄 Mocking 'expensive-api' tool, actual handler not called")
// Construct and return mock result directly
mockResult := mcp.NewTextResult("Mock response from middleware!")
return mockResult, nil // ← Handler is NOT called
}
// For other tools, continue to actual handler
return next(ctx, req)
}
}
```
**Key Point**: When you return a result directly, the actual tool handler is **bypassed**, saving execution time and resources.
#### Example 2: Intercepting and Enhancing Prompt Lists
```go
// PromptInterceptorMiddleware adds dynamic prompts to the list
func PromptInterceptorMiddleware(next mcp.HandlerFunc) mcp.HandlerFunc {
return func(ctx context.Context, req *mcp.JSONRPCRequest) (mcp.JSONRPCMessage, error) {
if req.Method != mcp.MethodPromptsList {
return next(ctx, req)
}
// Call original handler to get base prompt list
result, err := next(ctx, req)
if err != nil {
return nil, err
}
// Type assert and modify the result
if promptList, ok := result.(*mcp.ListPromptsResult); ok {
// Add a dynamic prompt generated by middleware
promptList.Prompts = append(promptList.Prompts, mcp.Prompt{
Name: "dynamic-prompt",
Description: "🎯 Dynamically added by middleware!",
Arguments: []mcp.PromptArgument{
{
Name: "topic",
Description: "Topic to discuss",
Required: false,
},
},
})
log.Printf("✅ Added 1 dynamic prompt to the list")
}
return result, err
}
}
```
#### Example 3: Intercepting prompts/get and Returning Cached Content
```go
func PromptCacheMiddleware(next mcp.HandlerFunc) mcp.HandlerFunc {
return func(ctx context.Context, req *mcp.JSONRPCRequest) (mcp.JSONRPCMessage, error) {
if req.Method != mcp.MethodPromptsGet {
return next(ctx, req)
}
// Parse prompt name
var getPromptReq mcp.GetPromptRequest
if params, ok := req.Params.(map[string]interface{}); ok {
if name, ok := params["name"].(string); ok {
getPromptReq.Params.Name = name
}
}
// Check cache for specific prompt
if getPromptReq.Params.Name == "cached-prompt" {
log.Printf("💾 Returning cached prompt content")
// Construct cached result directly
cachedResult := &mcp.GetPromptResult{
Description: "💾 Served from cache",
Messages: []mcp.PromptMessage{
{
Role: mcp.RoleUser,
Content: mcp.TextContent{
Type: "text",
Text: "This is cached content, loaded instantly!",
},
},
},
}
return cachedResult, nil // ← Handler is NOT called
}
// For uncached prompts, call actual handler
return next(ctx, req)
}
}
```
#### Example 4: Result Construction - All MCP Method Types
Middleware can construct results for any MCP method by returning the appropriate result type:
```go
func InterceptorMiddleware(next mcp.HandlerFunc) mcp.HandlerFunc {
return func(ctx context.Context, req *mcp.JSONRPCRequest) (mcp.JSONRPCMessage, error) {
switch req.Method {
case mcp.MethodToolsCall:
// Return CallToolResult
return &mcp.CallToolResult{
Content: []mcp.Content{
{Type: "text", Text: "Custom tool result"},
},
}, nil
case mcp.MethodPromptsGet:
// Return GetPromptResult
return &mcp.GetPromptResult{
Description: "Custom prompt",
Messages: []mcp.PromptMessage{
{
Role: mcp.RoleUser,
Content: mcp.TextContent{
Type: "text",
Text: "Custom prompt content",
},
},
},
}, nil
case mcp.MethodPromptsList:
// Return ListPromptsResult
return &mcp.ListPromptsResult{
Prompts: []mcp.Prompt{
{Name: "custom-prompt", Description: "Added by middleware"},
},
}, nil
case mcp.MethodResourcesRead:
// Return ReadResourceResult
return &mcp.ReadResourceResult{
Contents: []mcp.ResourceContents{
{
URI: "custom://resource",
MimeType: "text/plain",
Text: "Custom resource content",
},
},
}, nil
case mcp.MethodInitialize:
// Modify InitializeResult by calling handler first
result, err := next(ctx, req)
if err != nil {
return nil, err
}
if initResult, ok := result.(*mcp.InitializeResult); ok {
initResult.Instructions = "🎯 Enhanced by middleware!\n" + initResult.Instructions
}
return result, err
default:
return next(ctx, req)
}
}
}
```
#### Common Use Cases
**1. Tool Mocking/Testing**
```go
if toolName == "external-api" {
// Return mock data for testing
return mcp.NewTextResult("Mock API response"), nil
}
```
**2. Tool Caching**
```go
if cached := cache.Get(toolName + args); cached != nil {
return cached, nil // Return from cache
}
result, err := next(ctx, req) // Call handler
cache.Set(key, result) // Cache result
return result, err
```
**3. Graceful Degradation**
```go
if toolName == "unreliable-service" {
// Return fallback response instead of calling unstable service
return mcp.NewTextResult("Service unavailable, using fallback"), nil
}
```
**4. Access Control**
```go
if toolName == "admin-tool" {
session := mcp.ClientSessionFromContext(ctx)
if !isAdmin(session) {
result := mcp.NewTextResult("Access denied")
result.IsError = true
return result, nil
}
}
```
**5. Dynamic Prompt Loading**
```go
if req.Method == mcp.MethodPromptsGet {
// Load prompt from external source (DB, file, API)
promptContent := loadFromDatabase(promptName)
return &mcp.GetPromptResult{
Messages: []mcp.PromptMessage{{
Role: mcp.RoleUser,
Content: mcp.TextContent{Text: promptContent},
}},
}, nil
}
```
#### Complete Example
See [`examples/middleware/`](examples/middleware/) for a comprehensive demonstration including:
- ✅ Trace IDs
- ✅ Request logging
- ✅ Metrics collection
- ✅ Authorization checking
- ✅ Tool interception (mock/cache/degrade/block)
- ✅ Prompt interception (dynamic addition/caching)
- ✅ Result modification and construction
Run the example:
```bash
# Terminal 1: Start server
cd examples/middleware/server
go build && ./middleware-server
# Terminal 2: Run client
cd examples/middleware/client
go build && ./client
```
**Key Takeaways:**
- Middleware operates at the JSON-RPC layer, giving you full control over all MCP methods
- You can intercept **any** MCP request: tools, prompts, resources, initialize, ping, etc.
- Return results directly to bypass the actual handler (useful for mocking, caching, blocking)
- Call `next(ctx, req)` first, then modify the result for enhancement scenarios
- Use proper result types: `*CallToolResult`, `*GetPromptResult`, `*ListPromptsResult`, etc.
- Middleware executes in registration order (onion model)
## Copyright
The copyright notice pertaining to the Tencent code in this repo was previously in the name of “THL A29 Limited.” That entity has now been de-registered. You should treat all previously distributed copies of the code as if the copyright notice was in the name of “Tencent.”
连接信息
猜你喜欢
everything-claude-code
完整的 Claude Code 配置集合 - agents、skills、hooks、commands、rules、MCPs。来自 Anthropic...
Firecrawl
Firecrawl MCP Server 实现了网页抓取、爬虫和内容提取功能。
MarkItDown MCP
MarkItDown-MCP是一个轻量级服务器，用于将URI转换为Markdown。
illustrator-mcp-server
用于读取、操作和导出 Adobe Illustrator 设计数据的 MCP 服务器
rails-ai-context
只需安装一个 gem，即可将任何 Rails 应用程序转换为支持 AI 的代码库。通过 MCP 自动内省模型、路由、模式、任务并公开它们。
kakaocli
适用于AI代理的KakaoTalk CLI - 通过数据库和UI自动化读取和发送消息