Skip to content

server.md

Latest commit

 

History

History
1222 lines (1048 loc) · 46.3 KB

server.md

File metadata and controls

1222 lines (1048 loc) · 46.3 KB
Raw

Support for MCP server features

  1. Prompts
    1. Prompt message content
  2. Resources
    1. Binary resources
    2. Resource subscriptions
  3. Tools
    1. Tool result content
  4. List changed notifications
  5. Multi Round-Trip Requests
    1. Talking to legacy clients
    2. Example
  6. Cacheable list results
  7. Utilities
    1. Completion
    2. Logging
  8. Capabilities
    1. Capability inference
    2. Explicit capabilities
    3. Extensions
    4. Pagination

Prompts

MCP servers can provide LLM prompt templates (called simply prompts) to clients. Every prompt has a required name which identifies it, and a set of named arguments, which are strings.

Client-side: To list the server's prompts, use the ClientSession.Prompts iterator, or the lower-level ClientSession.ListPrompts (see pagination below). Set ClientOptions.PromptListChangedHandler to be notified of changes in the list of prompts.

Call ClientSession.GetPrompt to retrieve a prompt by name, providing arguments for expansion.

Server-side: Use Server.AddPrompt to add a prompt to the server along with its handler. The server will have the prompts capability if any prompt is added before the server is connected to a client, or if ServerOptions.HasPrompts is explicitly set. When a prompt is added, any clients already connected to the server will be notified via a notifications/prompts/list_changed notification.

func Example_prompts() {
	ctx := context.Background()

	promptHandler := func(ctx context.Context, req *mcp.GetPromptRequest) (*mcp.GetPromptResult, error) {
		return &mcp.GetPromptResult{
			Description: "Hi prompt",
			Messages: []*mcp.PromptMessage{
				{
					Role:    "user",
					Content: &mcp.TextContent{Text: "Say hi to " + req.Params.Arguments["name"]},
				},
			},
		}, nil
	}

	// Create a server with a single prompt.
	s := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, nil)
	prompt := &mcp.Prompt{
		Name: "greet",
		Arguments: []*mcp.PromptArgument{
			{
				Name:        "name",
				Description: "the name of the person to greet",
				Required:    true,
			},
		},
	}
	s.AddPrompt(prompt, promptHandler)

	// Create a client.
	c := mcp.NewClient(&mcp.Implementation{Name: "client", Version: "v0.0.1"}, nil)

	// Connect the server and client.
	t1, t2 := mcp.NewInMemoryTransports()
	if _, err := s.Connect(ctx, t1, nil); err != nil {
		log.Fatal(err)
	}
	cs, err := c.Connect(ctx, t2, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer cs.Close()

	// List the prompts.
	for p, err := range cs.Prompts(ctx, nil) {
		if err != nil {
			log.Fatal(err)
		}
		fmt.Println(p.Name)
	}

	// Get the prompt.
	res, err := cs.GetPrompt(ctx, &mcp.GetPromptParams{
		Name:      "greet",
		Arguments: map[string]string{"name": "Pat"},
	})
	if err != nil {
		log.Fatal(err)
	}
	for _, msg := range res.Messages {
		fmt.Println(msg.Role, msg.Content.(*mcp.TextContent).Text)
	}
	// Output:
	// greet
	// user Say hi to Pat
}

Prompt message content

A PromptMessage has two fields: a Role, either "user" or "assistant", and a single Content. That is the same interface a tool result uses (see Tool result content), so a prompt message can hold an image, audio, or a resource's contents as readily as text.

Media on its own says nothing about what the model should do with it, so send a text message alongside it. Embedding a resource puts its contents directly in the message, sparing the client a separate resources/read. Note that ResourceContents.URI is not checked against the resources the server has registered: it records where the contents came from, so use the URI of a real resource when there is one, as in the example below.

func Example_promptContent() {
	ctx := context.Background()

	const styleGuide = "- Prefer clarity over cleverness.\n"
	screenshotPNG := []byte("\x89PNG\r\n\x1a\n")

	s := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, nil)

	s.AddResource(&mcp.Resource{URI: "doc://style-guide", MIMEType: "text/markdown"},
		func(context.Context, *mcp.ReadResourceRequest) (*mcp.ReadResourceResult, error) {
			return &mcp.ReadResourceResult{
				Contents: []*mcp.ResourceContents{{
					URI:      "doc://style-guide",
					MIMEType: "text/markdown",
					Text:     styleGuide,
				}},
			}, nil
		})

	s.AddPrompt(&mcp.Prompt{Name: "review_screenshot"},
		func(context.Context, *mcp.GetPromptRequest) (*mcp.GetPromptResult, error) {
			return &mcp.GetPromptResult{
				Messages: []*mcp.PromptMessage{
					{Role: "user", Content: &mcp.EmbeddedResource{
						Resource: &mcp.ResourceContents{
							URI:      "doc://style-guide",
							MIMEType: "text/markdown",
							Text:     styleGuide,
						},
					}},
					{Role: "user", Content: &mcp.ImageContent{Data: screenshotPNG, MIMEType: "image/png"}},
					{Role: "user", Content: &mcp.TextContent{Text: "Review the screenshot against the style guide."}},
				},
			}, nil
		})

	c := mcp.NewClient(&mcp.Implementation{Name: "client", Version: "v0.0.1"}, nil)
	t1, t2 := mcp.NewInMemoryTransports()
	if _, err := s.Connect(ctx, t1, nil); err != nil {
		log.Fatal(err)
	}
	cs, err := c.Connect(ctx, t2, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer cs.Close()

	res, err := cs.GetPrompt(ctx, &mcp.GetPromptParams{Name: "review_screenshot"})
	if err != nil {
		log.Fatal(err)
	}
	for _, msg := range res.Messages {
		switch content := msg.Content.(type) {
		case *mcp.TextContent:
			fmt.Println(msg.Role, "text:", content.Text)
		case *mcp.ImageContent:
			fmt.Printf("%s image: %s, %d bytes\n", msg.Role, content.MIMEType, len(content.Data))
		case *mcp.EmbeddedResource:
			fmt.Printf("%s embedded resource: %s (%s)\n", msg.Role, content.Resource.URI, content.Resource.MIMEType)
		}
	}
	// Output:
	// user embedded resource: doc://style-guide (text/markdown)
	// user image: image/png, 8 bytes
	// user text: Review the screenshot against the style guide.
}

Resources

In MCP terms, a resource is some data referenced by a URI. MCP servers can serve resources to clients. They can register resources individually, or register a resource template that uses a URI pattern to describe a collection of resources.

Client-side: Call ClientSession.ReadResource to read a resource. The SDK ensures that a read succeeds only if the URI matches a registered resource exactly, or matches the URI pattern of a resource template.

To list a server's resources and resource templates, use the ClientSession.Resources and ClientSession.ResourceTemplates iterators, or the lower-level ListXXX calls (see pagination). Set ClientOptions.ResourceListChangedHandler to be notified of changes in the lists of resources or resource templates.

Clients can be notified when the contents of a resource changes by subscribing to the resource's URI. Call ClientSession.Subscribe to subscribe to a resource and ClientSession.Unsubscribe to unsubscribe. Set ClientOptions.ResourceUpdatedHandler to be notified of changes to subscribed resources. On 2026-07-28 and later sessions the SDK delivers these notifications over a subscriptions/listen stream instead of the legacy resources/subscribe RPC; see Subscriptions (subscriptions/listen) for the wire-level details.

Server-side: Use Server.AddResource or Server.AddResourceTemplate to add a resource or resource template to the server along with its handler. A ResourceHandler maps a URI to the contents of a resource, which can include text, binary data, or both. If AddResource or AddResourceTemplate is called before a server is connected, the server will have the resources capability. The server will have the resources capability if any resource or resource template is added before the server is connected to a client, or if ServerOptions.HasResources is explicitly set. When a prompt is added, any clients already connected to the server will be notified via a notifications/resources/list_changed notification.

func Example_resources() {
	ctx := context.Background()

	resources := map[string]string{
		"file:///a":     "a",
		"file:///dir/x": "x",
		"file:///dir/y": "y",
	}

	handler := func(_ context.Context, req *mcp.ReadResourceRequest) (*mcp.ReadResourceResult, error) {
		uri := req.Params.URI
		c, ok := resources[uri]
		if !ok {
			return nil, mcp.ResourceNotFoundError(uri)
		}
		return &mcp.ReadResourceResult{
			Contents: []*mcp.ResourceContents{{URI: uri, Text: c}},
		}, nil
	}

	// Create a server with a single resource.
	s := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, nil)
	s.AddResource(&mcp.Resource{URI: "file:///a"}, handler)
	s.AddResourceTemplate(&mcp.ResourceTemplate{URITemplate: "file:///dir/{f}"}, handler)

	// Create a client.
	c := mcp.NewClient(&mcp.Implementation{Name: "client", Version: "v0.0.1"}, nil)

	// Connect the server and client.
	t1, t2 := mcp.NewInMemoryTransports()
	if _, err := s.Connect(ctx, t1, nil); err != nil {
		log.Fatal(err)
	}
	cs, err := c.Connect(ctx, t2, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer cs.Close()

	// List resources and resource templates.
	for r, err := range cs.Resources(ctx, nil) {
		if err != nil {
			log.Fatal(err)
		}
		fmt.Println(r.URI)
	}
	for r, err := range cs.ResourceTemplates(ctx, nil) {
		if err != nil {
			log.Fatal(err)
		}
		fmt.Println(r.URITemplate)
	}

	// Read resources.
	for _, path := range []string{"a", "dir/x", "b"} {
		res, err := cs.ReadResource(ctx, &mcp.ReadResourceParams{URI: "file:///" + path})
		if err != nil {
			fmt.Println(err)
		} else {
			fmt.Println(res.Contents[0].Text)
		}
	}
	// Output:
	// file:///a
	// file:///dir/{f}
	// a
	// x
	// calling "resources/read": Resource not found
}

Binary resources

A ResourceContents carries its data in either Text or Blob. Binary data goes in Blob as plain bytes — the SDK base64-encodes it on the wire — and leaves Text empty, so a client tells the two apart by which field is populated.

func Example_binaryResource() {
	ctx := context.Background()

	s := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, nil)
	s.AddResource(&mcp.Resource{URI: "file:///logo.png", MIMEType: "image/png"},
		func(context.Context, *mcp.ReadResourceRequest) (*mcp.ReadResourceResult, error) {
			return &mcp.ReadResourceResult{
				Contents: []*mcp.ResourceContents{{
					URI:      "file:///logo.png",
					MIMEType: "image/png",
					Blob:     []byte("\x89PNG\r\n\x1a\n"),
				}},
			}, nil
		})

	c := mcp.NewClient(&mcp.Implementation{Name: "client", Version: "v0.0.1"}, nil)
	t1, t2 := mcp.NewInMemoryTransports()
	if _, err := s.Connect(ctx, t1, nil); err != nil {
		log.Fatal(err)
	}
	cs, err := c.Connect(ctx, t2, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer cs.Close()

	res, err := cs.ReadResource(ctx, &mcp.ReadResourceParams{URI: "file:///logo.png"})
	if err != nil {
		log.Fatal(err)
	}
	contents := res.Contents[0]
	fmt.Printf("%s: %d bytes of %s, text is %q\n",
		contents.URI, len(contents.Blob), contents.MIMEType, contents.Text)

	// Output:
	// file:///logo.png: 8 bytes of image/png, text is ""
}

Resource subscriptions

A client interested in changes to one resource subscribes to its URI with ClientSession.Subscribe and stops with ClientSession.Unsubscribe. Set ServerOptions.SubscribeHandler and UnsubscribeHandler to track who is listening; setting either one gives the server the resources.subscribe capability.

Publish a change with Server.ResourceUpdated, which notifies every session subscribed to that URI. The notification carries only the URI, not the new contents, so a client that wants them reads the resource again.

func Example_resourceSubscription() {
	ctx := context.Background()

	config := "theme=light\n"

	updated := make(chan string, 1)
	c := mcp.NewClient(&mcp.Implementation{Name: "client", Version: "v0.0.1"}, &mcp.ClientOptions{
		// The notification carries only the URI, so read the resource again to
		// see what it now holds.
		ResourceUpdatedHandler: func(context.Context, *mcp.ResourceUpdatedNotificationRequest) {
			updated <- "config://app"
		},
	})

	subscribed := make(chan string, 2)
	s := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, &mcp.ServerOptions{
		SubscribeHandler: func(_ context.Context, req *mcp.SubscribeRequest) error {
			subscribed <- "subscribed to " + req.Params.URI
			return nil
		},
		UnsubscribeHandler: func(_ context.Context, req *mcp.UnsubscribeRequest) error {
			subscribed <- "unsubscribed from " + req.Params.URI
			return nil
		},
	})
	s.AddResource(&mcp.Resource{URI: "config://app"},
		func(context.Context, *mcp.ReadResourceRequest) (*mcp.ReadResourceResult, error) {
			return &mcp.ReadResourceResult{
				Contents: []*mcp.ResourceContents{{URI: "config://app", Text: config}},
			}, nil
		})

	t1, t2 := mcp.NewInMemoryTransports()
	if _, err := s.Connect(ctx, t1, nil); err != nil {
		log.Fatal(err)
	}
	cs, err := c.Connect(ctx, t2, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer cs.Close()

	if err := cs.Subscribe(ctx, &mcp.SubscribeParams{URI: "config://app"}); err != nil {
		log.Fatal(err)
	}
	fmt.Println(<-subscribed)

	config = "theme=dark\n"
	if err := s.ResourceUpdated(ctx, &mcp.ResourceUpdatedNotificationParams{URI: "config://app"}); err != nil {
		log.Fatal(err)
	}
	uri := <-updated

	res, err := cs.ReadResource(ctx, &mcp.ReadResourceParams{URI: uri})
	if err != nil {
		log.Fatal(err)
	}
	fmt.Print("re-read: ", res.Contents[0].Text)

	if err := cs.Unsubscribe(ctx, &mcp.UnsubscribeParams{URI: "config://app"}); err != nil {
		log.Fatal(err)
	}
	fmt.Println(<-subscribed)

	// Output:
	// subscribed to config://app
	// re-read: theme=dark
	// unsubscribed from config://app
}

Tools

MCP servers can provide tools to allow clients to interact with external systems or functionality. Tools are effectively remote function calls, and the Go SDK provides mechanisms to bind them to ordinary Go functions.

Client-side: To list the server's tools, use the ClientSession.Tools iterator, or the lower-level ClientSession.ListTools (see pagination). Set ClientOptions.ToolListChangedHandler to be notified of changes in the list of tools.

To call a tool, use ClientSession.CallTool with CallToolParams holding the name and arguments of the tool to call.

res, err := session.CallTool(ctx, &mcp.CallToolParams{
	Name:      "my_tool",
	Arguments: map[string]any{"name": "user"},
})

Arguments may be any value that can be marshaled to JSON.

Server-side: the basic API for adding a tool is symmetrical with the API for prompts or resources: Server.AddTool adds a Tool to the server along with its ToolHandler to handle it. The server will have the tools capability if any tool is added before the server is connected to a client, or if ServerOptions.HasTools is explicitly set. When a tool is added, any clients already connected to the server will be notified via a notifications/tools/list_changed notification.

However, the Server.AddTool API leaves it to the user to implement the tool handler correctly according to the spec, providing very little out of the box. In order to implement a tool, the user must do all of the following:

  • Provide a tool input and output schema.
  • Validate the tool arguments against its input schema.
  • Unmarshal the input schema into a Go value
  • Execute the tool logic.
  • Marshal the tool's structured output (if any) to JSON, and store it in the result's StructuredContent field as well as the unstructured Content field.
  • Validate that output JSON against the tool's output schema.
  • If any tool errors occurred, pack them into the unstructured content and set IsError to true.

For this reason, the SDK provides a generic AddTool function that handles this for you. It can bind a tool to any function with the following shape:

func(_ context.Context, request *CallToolRequest, input In) (result *CallToolResult, output Out, _ error)

This is like a ToolHandler, but with an extra arbitrary In input parameter, and Out output parameter.

Such a function can then be bound to the server using AddTool:

mcp.AddTool(server, &mcp.Tool{Name: "my_tool"}, handler)

This does the following automatically:

  • If Tool.InputSchema is unset, the input schema is inferred from the In type, which must be a struct or map.
  • If Tool.OutputSchema is unset and the Out type is not any, the output schema is inferred from the Out type. Per SEP-2106, Out may be any Go type whose inferred schema is a valid JSON Schema (struct, map, slice, primitive, etc.).
  • Optional jsonschema struct tags provide argument and output descriptions.
  • Tool arguments are validated against the input schema.
  • Tool arguments are marshaled into the In value.
  • Tool output (the Out value) is marshaled into the result's StructuredContent, as well as the unstructured Content.
  • Output is validated against the tool's output schema.
  • If an ordinary error is returned, it is stored int the CallToolResult and IsError is set to true.

In fact, under ordinary circumstances, the user can ignore CallToolRequest and CallToolResult.

For a more realistic example, consider a tool that retrieves the weather:

type WeatherInput struct {
	Location Location `json:"location" jsonschema:"user location"`
	Days     int      `json:"days" jsonschema:"number of days to forecast"`
}

type WeatherOutput struct {
	Summary       string      `json:"summary" jsonschema:"a summary of the weather forecast"`
	Confidence    Probability `json:"confidence" jsonschema:"confidence, between 0 and 1"`
	AsOf          time.Time   `json:"asOf" jsonschema:"the time the weather was computed"`
	DailyForecast []Forecast  `json:"dailyForecast" jsonschema:"the daily forecast"`
	Source        string      `json:"source,omitempty" jsonschema:"the organization providing the weather forecast"`
}

func WeatherTool(ctx context.Context, req *mcp.CallToolRequest, in WeatherInput) (*mcp.CallToolResult, WeatherOutput, error) {
	perfectWeather := WeatherOutput{
		Summary:    "perfect",
		Confidence: 1.0,
		AsOf:       time.Now(),
	}
	for range in.Days {
		perfectWeather.DailyForecast = append(perfectWeather.DailyForecast, Forecast{
			Forecast: "another perfect day",
			Type:     Sunny,
			Rain:     0.0,
			High:     72.0,
			Low:      72.0,
		})
	}
	return nil, perfectWeather, nil
}

In this case, we want to customize part of the inferred schema, though we can still infer the rest. Since we want to control the inference ourselves, we set the Tool.InputSchema explicitly:

// Distinguished Go types allow custom schemas to be reused during inference.
customSchemas := map[reflect.Type]*jsonschema.Schema{
	reflect.TypeFor[Probability](): {Type: "number", Minimum: jsonschema.Ptr(0.0), Maximum: jsonschema.Ptr(1.0)},
	reflect.TypeFor[WeatherType](): {Type: "string", Enum: []any{Sunny, PartlyCloudy, Cloudy, Rainy, Snowy}},
}
opts := &jsonschema.ForOptions{TypeSchemas: customSchemas}
in, err := jsonschema.For[WeatherInput](opts)
if err != nil {
	log.Fatal(err)
}

// Furthermore, we can tweak the inferred schema, in this case limiting
// forecasts to 0-10 days.
daysSchema := in.Properties["days"]
daysSchema.Minimum = jsonschema.Ptr(0.0)
daysSchema.Maximum = jsonschema.Ptr(10.0)

// Output schema inference can reuse our custom schemas from input inference.
out, err := jsonschema.For[WeatherOutput](opts)
if err != nil {
	log.Fatal(err)
}

// Now add our tool to a server. Since we've customized the schemas, we need
// to override the default schema inference.
server := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, nil)
mcp.AddTool(server, &mcp.Tool{
	Name:         "weather",
	InputSchema:  in,
	OutputSchema: out,
}, WeatherTool)

See mcp/tool_example_test.go for the full example, or examples/server/toolschemas for more examples of customizing tool schemas.

Tool result content

Alongside its structured output, a tool result carries a list of Content blocks in CallToolResult.Content, and may mix as many kinds as it needs:

Type Carries
TextContent plain text
ImageContent image data, with a MIME type
AudioContent audio data, with a MIME type
EmbeddedResource the contents of a resource, inline
ResourceLink a reference to a resource the client can read separately

ImageContent.Data and AudioContent.Data are plain []byte; the SDK base64-encodes them on the wire, so handlers never encode by hand, and base64 carried over from another MCP SDK must be decoded before it is assigned. An EmbeddedResource wraps a ResourceContents, which holds either Text or, for binary data, Blob. Prefer a ResourceLink when the client may not need the contents, since an embedded resource is transferred whether it is used or not.

When a handler bound with AddTool leaves Content unset, the SDK fills it with the JSON encoding of the output value; set Content explicitly, as below, to return anything else. StructuredContent is still populated from the output value either way.

func ExampleAddTool_contentTypes() {
	chartPNG := []byte("\x89PNG\r\n\x1a\n")
	summaryWAV := []byte("RIFF....WAVE")

	server := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, nil)
	mcp.AddTool(server, &mcp.Tool{Name: "report"}, func(context.Context, *mcp.CallToolRequest, struct{}) (*mcp.CallToolResult, any, error) {
		return &mcp.CallToolResult{
			Content: []mcp.Content{
				&mcp.TextContent{Text: "Q3 revenue by region."},
				&mcp.ImageContent{Data: chartPNG, MIMEType: "image/png"},
				&mcp.AudioContent{Data: summaryWAV, MIMEType: "audio/wav"},
				// An embedded resource inlines the contents, so the client
				// needs no follow-up resources/read. Binary contents go in
				// ResourceContents.Blob instead of Text.
				&mcp.EmbeddedResource{
					Resource: &mcp.ResourceContents{
						URI:      "file:///reports/q3.csv",
						MIMEType: "text/csv",
						Text:     "region,revenue\nEMEA,42\n",
					},
				},
				// A resource link only points at a resource. Prefer it when the
				// client may not need the contents, since an embedded resource
				// is transferred either way.
				&mcp.ResourceLink{URI: "file:///reports/q3.pdf", MIMEType: "application/pdf"},
			},
		}, nil, nil
	})

	ctx := context.Background()
	session, err := connect(ctx, server)
	if err != nil {
		log.Fatal(err)
	}
	defer session.Close()

	res, err := session.CallTool(ctx, &mcp.CallToolParams{Name: "report"})
	if err != nil {
		log.Fatal(err)
	}

	// Clients type-switch over the content they receive.
	for _, content := range res.Content {
		switch c := content.(type) {
		case *mcp.TextContent:
			fmt.Println("text:", c.Text)
		case *mcp.ImageContent:
			fmt.Printf("image: %s, %d bytes\n", c.MIMEType, len(c.Data))
		case *mcp.AudioContent:
			fmt.Printf("audio: %s, %d bytes\n", c.MIMEType, len(c.Data))
		case *mcp.EmbeddedResource:
			fmt.Printf("embedded resource: %s (%s)\n", c.Resource.URI, c.Resource.MIMEType)
		case *mcp.ResourceLink:
			fmt.Printf("resource link: %s (%s)\n", c.URI, c.MIMEType)
		}
	}
	// Output:
	// text: Q3 revenue by region.
	// image: image/png, 8 bytes
	// audio: audio/wav, 12 bytes
	// embedded resource: file:///reports/q3.csv (text/csv)
	// resource link: file:///reports/q3.pdf (application/pdf)
}

ToolUseContent and ToolResultContent also implement Content, but are only valid in sampling messages, not in tool results.

Stateless server deployments: Some deployments create a new Server for each incoming request, re-registering tools every time. To avoid repeated schema generation, create a SchemaCache and share it across server instances:

var schemaCache = mcp.NewSchemaCache() // create once at startup

func handleRequest(w http.ResponseWriter, r *http.Request) {
    s := mcp.NewServer(impl, &mcp.ServerOptions{SchemaCache: schemaCache})
    mcp.AddTool(s, myTool, myHandler)
    // ...
}

List changed notifications

Adding or removing a feature on a connected server sends the matching notifications/*/list_changed to every client, which dispatches it to ClientOptions.ToolListChangedHandler, PromptListChangedHandler, or ResourceListChangedHandler. A notification reports only that the list changed, so a client that needs the new contents lists them again.

These notifications ride on capabilities, and capabilities are inferred from what is registered before Server.Connect (see Capabilities). A server that registers its features afterwards, as the example below does, has to declare ServerOptions.HasTools, HasPrompts, or HasResources itself, or the client is never told.

func Example_listChanged() {
	ctx := context.Background()

	changed := make(chan string, 3)
	c := mcp.NewClient(&mcp.Implementation{Name: "client", Version: "v0.0.1"}, &mcp.ClientOptions{
		ToolListChangedHandler: func(context.Context, *mcp.ToolListChangedRequest) {
			changed <- "tools"
		},
		PromptListChangedHandler: func(context.Context, *mcp.PromptListChangedRequest) {
			changed <- "prompts"
		},
		ResourceListChangedHandler: func(context.Context, *mcp.ResourceListChangedRequest) {
			changed <- "resources"
		},
	})

	// Nothing is registered before Connect, so the capabilities that carry these
	// notifications have to be declared explicitly.
	s := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, &mcp.ServerOptions{
		HasTools:     true,
		HasPrompts:   true,
		HasResources: true,
	})

	t1, t2 := mcp.NewInMemoryTransports()
	if _, err := s.Connect(ctx, t1, nil); err != nil {
		log.Fatal(err)
	}
	cs, err := c.Connect(ctx, t2, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer cs.Close()

	mcp.AddTool(s, &mcp.Tool{Name: "greet"}, func(context.Context, *mcp.CallToolRequest, struct{}) (*mcp.CallToolResult, any, error) {
		return &mcp.CallToolResult{}, nil, nil
	})
	fmt.Println("changed:", <-changed)

	s.AddPrompt(&mcp.Prompt{Name: "review"}, func(context.Context, *mcp.GetPromptRequest) (*mcp.GetPromptResult, error) {
		return &mcp.GetPromptResult{}, nil
	})
	fmt.Println("changed:", <-changed)

	s.AddResource(&mcp.Resource{URI: "file:///a"}, func(context.Context, *mcp.ReadResourceRequest) (*mcp.ReadResourceResult, error) {
		return &mcp.ReadResourceResult{}, nil
	})
	fmt.Println("changed:", <-changed)

	// Removing a feature notifies the client just as adding one does.
	s.RemoveTools("greet")
	fmt.Println("changed:", <-changed)

	// Output:
	// changed: tools
	// changed: prompts
	// changed: resources
	// changed: tools
}

Multi Round-Trip Requests

SEP-2322 defines a new pattern for server-to-client requests (sampling, elicitation, roots). Instead of issuing a fresh JSON-RPC request mid-flight, the server returns an InputRequiredResult from its in-flight handler — the InputRequests field of CallToolResult, GetPromptResult, or ReadResourceResult carries the requests for additional information. The client responds by retrying the original call with InputResponses set.

The SDK supports this pattern from both sides without requiring callers to choose:

  • Returning input requests from a handler. Set InputRequests on the result you return. Leave Content / StructuredContent empty — returning both at once is a server bug and the SDK returns -32603 InternalError.
  • Calling the legacy APIs. ServerSession.Elicit, ServerSession.CreateMessage(WithTools), and ServerSession.ListRoots remain available and work for both new and legacy clients.

Talking to legacy clients

The server installs serverMultiRoundTripMiddleware automatically. For clients on a protocol version earlier than 2026-07-28, the middleware intercepts any InputRequiredResult your handler returns, fulfils each input request itself by calling the legacy server-initiated APIs (Elicit, CreateMessage, ListRoots), and re-invokes your handler exactly once with the responses already populated. This means a handler written in the MRTR style works against both old and new clients without code changes.

Example

The following example shows a "greet" tool whose handler asks the user for their name via elicitation before producing the final greeting. The handler runs twice — once to issue the elicitation, once to consume the response — but the client call site sees a single CallTool returning the final result, because the SDK's MRTR middleware handles the round trip on either side of the wire (depending on the negotiated protocol version).

// Example_mrtr demonstrates the [Multi Round-Trip Requests] pattern
// (SEP-2322). A tool handler signals "I need more information from the user"
// by returning a [mcp.CallToolResult] whose `InputRequests` field carries the
// requests for the additional information. Each request can be an
// [mcp.ElicitParams] (ask the user a question),
// [mcp.CreateMessageParams] (sample from the client's LLM), or
// [mcp.ListRootsParams] (list the client's roots).
//
// On protocol version `2026-07-28` and later, the client's
// `clientMultiRoundTripMiddleware` fulfils each input request from the
// configured handler and retries the original call transparently. On earlier
// protocol versions, the server's `serverMultiRoundTripMiddleware` performs
// the equivalent dance from the server side by calling the legacy
// server-to-client API (`ServerSession.Elicit`, `CreateMessage`, or
// `ListRoots`) and re-invoking the handler with the response. Either way,
// the user-facing client call site sees only the final result.
//
// [Multi Round-Trip Requests]: https://modelcontextprotocol.io/specification/draft/basic/patterns#multi-round-trip-requests
func Example_mrtr() {
	ctx := context.Background()

	// Server: a "greet" tool that asks the user for their name before
	// returning a greeting. The handler runs twice per logical call:
	// once to issue the elicitation, once to consume the response.
	s := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, nil)
	mcp.AddTool(s, &mcp.Tool{Name: "greet"}, func(_ context.Context, req *mcp.CallToolRequest, _ struct{}) (*mcp.CallToolResult, any, error) {
		if len(req.Params.InputResponses) == 0 {
			// First call: ask the user for their name. The map key
			// ("user_name") is a label the handler chooses; the client must
			// echo it back in InputResponses. The wire-level method name
			// ("elicitation/create") is derived from the value's Go type.
			return &mcp.CallToolResult{
				InputRequests: mcp.InputRequestMap{
					"user_name": &mcp.ElicitParams{
						Message: "What's your name?",
						RequestedSchema: &jsonschema.Schema{
							Type: "object",
							Properties: map[string]*jsonschema.Schema{
								"name": {Type: "string"},
							},
						},
					},
				},
				// RequestState is an opaque token the client echoes back on
				// the retry, letting the handler resume its work without
				// per-session storage.
				RequestState: "step=1",
			}, nil, nil
		}
		// Retry: read the elicitation response and produce the final greeting.
		name := req.Params.InputResponses["user_name"].(*mcp.ElicitResult).Content["name"].(string)
		return &mcp.CallToolResult{
			Content: []mcp.Content{&mcp.TextContent{Text: "Hello " + name}},
		}, nil, nil
	})

	// Client: declares an elicitation handler. The SDK's MRTR middleware
	// uses it to fulfil any input request the tool handler returns.
	c := mcp.NewClient(&mcp.Implementation{Name: "client", Version: "v0.0.1"}, &mcp.ClientOptions{
		ElicitationHandler: func(_ context.Context, _ *mcp.ElicitRequest) (*mcp.ElicitResult, error) {
			return &mcp.ElicitResult{Action: "accept", Content: map[string]any{"name": "MCP Go"}}, nil
		},
	})

	ct, st := mcp.NewInMemoryTransports()
	if _, err := s.Connect(ctx, st, nil); err != nil {
		log.Fatal(err)
	}
	cs, err := c.Connect(ctx, ct, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer cs.Close()

	// Single call site. The MRTR round trip is invisible to the caller.
	res, err := cs.CallTool(ctx, &mcp.CallToolParams{Name: "greet"})
	if err != nil {
		log.Fatal(err)
	}
	fmt.Println(res.Content[0].(*mcp.TextContent).Text)
}

Cacheable list results

SEP-2549 adds ttlMs and cacheScope fields to the results of tools/list, prompts/list, resources/list, resources/templates/list, and resources/read. Both fields complement existing list results: clients use ttlMs as a freshness hint to reduce polling, and cacheScope ("public" or "private") controls whether shared intermediaries may cache the response.

Server-side defaults. After paginating, the SDK sets CacheScope = "public" and leaves TTLMs = 0 (which the client cache treats as "immediately stale"). A handler that wants its responses cached must set TTLMs explicitly on the returned result.

mcp.AddTool(server, &mcp.Tool{Name: "expensive"}, func(...) (*mcp.CallToolResult, any, error) {
    // ...
})
// Override the default for tools/list:
server.AddSendingMiddleware(func(next mcp.MethodHandler) mcp.MethodHandler {
    return func(ctx context.Context, method string, req mcp.Request) (mcp.Result, error) {
        res, err := next(ctx, method, req)
        if lr, ok := res.(*mcp.ListToolsResult); ok {
            lr.TTLMs = 30_000 // 30s cache freshness hint
        }
        return res, err
    }
})

Utilities

Completion

To support the completion capability, the server needs a completion handler.

Client-side: completion is called using the ClientSession.Complete method.

Server-side: completion is enabled by setting ServerOptions.CompletionHandler. If this field is set to a non-nil value, the server will advertise the completions server capability, and use this handler to respond to completion requests.

myCompletionHandler := func(_ context.Context, req *mcp.CompleteRequest) (*mcp.CompleteResult, error) {
	// In a real application, you'd implement actual completion logic here.
	// For this example, we return a fixed set of suggestions.
	var suggestions []string
	switch req.Params.Ref.Type {
	case "ref/prompt":
		suggestions = []string{"suggestion1", "suggestion2", "suggestion3"}
	case "ref/resource":
		suggestions = []string{"suggestion4", "suggestion5", "suggestion6"}
	default:
		return nil, fmt.Errorf("unrecognized content type %s", req.Params.Ref.Type)
	}

	return &mcp.CompleteResult{
		Completion: mcp.CompletionResultDetails{
			HasMore: false,
			Total:   len(suggestions),
			Values:  suggestions,
		},
	}, nil
}

// Create the MCP Server instance and assign the handler.
// No server running, just showing the configuration.
_ = mcp.NewServer(&mcp.Implementation{Name: "server"}, &mcp.ServerOptions{
	CompletionHandler: myCompletionHandler,
})

Logging

Note: The logging feature is deprecated as of protocol version 2026-07-28 (SEP-2577). It remains fully functional during the deprecation window (at least twelve months). The SDK continues to support logging for compatibility. Servers should migrate to stderr logging (for STDIO transports) or OpenTelemetry.

MCP servers can send logging messages to MCP clients. (This form of logging is distinct from server-side logging, where the server produces logs that remain server-side, for use by server maintainers.)

Server-side: The minimum log level is part of the server state. For stateful sessions, there is no default log level: no log messages will be sent until the client calls SetLevel (see below). For stateful sessions, the level defaults to "info".

ServerSession.Log is the low-level way for servers to log to clients. It sends a logging notification to the client if the level of the message is at least the minimum log level.

For a simpler API, use NewLoggingHandler to obtain a slog.Handler. By setting LoggingHandlerOptions.MinInterval, the handler can be rate-limited to avoid spamming clients with too many messages.

Servers always report the logging capability.

Client-side: Set ClientOptions.LoggingMessageHandler to receive log messages.

Call ClientSession.SetLevel to change the log level for a session.

func Example_logging() {
	ctx := context.Background()

	// Create a server.
	s := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "v0.0.1"}, nil)

	// Create a client that displays log messages.
	done := make(chan struct{}) // solely for the example
	var nmsgs atomic.Int32
	c := mcp.NewClient(
		&mcp.Implementation{Name: "client", Version: "v0.0.1"},
		&mcp.ClientOptions{
			LoggingMessageHandler: func(_ context.Context, r *mcp.LoggingMessageRequest) {
				m := r.Params.Data.(map[string]any)
				fmt.Println(m["msg"], m["value"])
				if nmsgs.Add(1) == 2 { // number depends on logger calls below
					close(done)
				}
			},
		})

	// Connect the server and client.
	t1, t2 := mcp.NewInMemoryTransports()
	ss, err := s.Connect(ctx, t1, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer ss.Close()
	cs, err := c.Connect(ctx, t2, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer cs.Close()

	// Set the minimum log level to "info".
	if err := cs.SetLoggingLevel(ctx, &mcp.SetLoggingLevelParams{Level: "info"}); err != nil {
		log.Fatal(err)
	}

	// Get a slog.Logger for the server session.
	logger := slog.New(mcp.NewLoggingHandler(ss, nil))

	// Log some things.
	logger.Info("info shows up", "value", 1)
	logger.Debug("debug doesn't show up", "value", 2)
	logger.Warn("warn shows up", "value", 3)

	// Wait for them to arrive on the client.
	// In a real application, the log messages would appear asynchronously
	// while other work was happening.
	<-done

	// Output:
	// info shows up 1
	// warn shows up 3
}

Capabilities

Server capabilities are advertised to clients during the initialization handshake. By default, the SDK advertises only the logging capability. Additional capabilities are automatically added when features are registered (e.g., adding a tool adds the tools capability).

Capability inference

When features such as tools, prompts, or resources are added to the server (e.g., via Server.AddTool), their capability is automatically inferred, with default value {listChanged:true}. Similarly, if the ServerOptions.SubscribeHandler or ServerOptions.CompletionHandler are set, the corresponding capability is added.

Explicit capabilities

To explicitly declare capabilities, or to override the default inferred capability, set ServerOptions.Capabilities. This sets the default server capabilities, before any capabilities are added based on configured handlers. If a capability is already present as a field in Capabilities, adding a feature or handler will not change its configuration.

This allows you to:

  • Disable default capabilities: Pass an empty &ServerCapabilities{} to disable all defaults, including logging.
  • Disable listChanged notifications: Set ListChanged: false on a capability to prevent the server from sending list-changed notifications when features are added or removed.
  • Pre-declare capabilities: Declare capabilities before features are registered, useful for servers that load features dynamically.
// Disable listChanged notifications for tools
server := mcp.NewServer(impl, &mcp.ServerOptions{
    Capabilities: &mcp.ServerCapabilities{
        Logging: &mcp.LoggingCapabilities{},
        Tools:   &mcp.ToolCapabilities{ListChanged: false},
    },
})

Deprecated: The HasPrompts, HasResources, and HasTools fields on ServerOptions are deprecated. Use Capabilities instead.

Extensions

SEP-2133 adds an extensions map to ServerCapabilities so that optional capabilities outside the core protocol can be declared on the wire. Keys are namespaced as "{vendor-prefix}/{extension-name}"; values are per-extension settings objects.

Pagination

Server-side feature lists may be paginated, using cursors. The SDK supports this by default.

Client-side: The ClientSession provides methods returning iterators for each feature type. These iterators are an iter.Seq2[Feature, error], where the error value indicates whether page retrieval failed.

The ClientSession also exposes ListXXX methods for fine-grained control over pagination.

Server-side: pagination is on by default, so in general nothing is required server-side. However, you may use ServerOptions.PageSize to customize the page size.