Go SDK: Design

[findleyr]

Pre-submission Checklist

• I have verified this would not be more appropriate as a feature request in a specific repository
• I have searched existing discussions to avoid duplicates

Your Idea

Hi everyone 👋! This is a design discussion for the Go MCP SDK, following up on /discussions/224, where we (the Go team) volunteered to help contribute an official SDK.

In the original discussion, many people expressed support for a Go SDK, which is fantastic. With so much interest, we felt that it was more important than ever to start with a concrete design; otherwise we’d risk a lot of churn as we all worked out the APIs incrementally.

We analyzed a number of SDKs in Go and other languages to arrive at the design below. Wherever possible, we tried to align with the very successful mark3labs/mcp-go. However, as you’ll see, we diverged in a number of ways in order to keep the APIs minimal and to be cautious about future spec evolution. We address the differences between this design and mcp-go throughout the document.

Our goal is to make sure that there is a good official SDK for Go. We want your feedback. If discussion reveals that some of the APIs below aren’t right, we’ll change them.

To achieve this, let’s proceed as follows:

• Start a comment thread here to discuss aspects of the design. As much as possible, try to keep threads focused on one area. Prefer to reuse existing threads rather than creating new ones.
• If compelling arguments emerge that the design needs to be changed, we’ll update it. We’ll also keep a log of any such changes.
• We hope that we can reach consensus within a reasonable amount of time (a couple of weeks, perhaps?). Once that happens, we’ll finalize the design and move to the execution phase, where we create github.com/modelcontextprotocol/go-sdk.

Thanks!
-Rob

Changelog

• 2025-05-22: Updated middleware per discussion; updated logging to include a general logging handler in addition to a slog.logger. changes

Go SDK Design

This document discusses the design of a Go SDK for the model context protocol. The golang.org/x/tools/internal/mcp package contains a prototype that we built to explore the MCP design space. Many of the ideas there are present in this document. However, we have diverged from and expanded on the APIs of that prototype, and this document should be considered canonical.

Similarities and differences with mark3labs/mcp-go (and others)

The most popular unofficial MCP SDK for Go is mark3labs/mcp-go. As of this writing, it is imported by over 400 packages that span over 200 modules.

We admire mcp-go, and where possible tried to align with its design. However, the APIs here diverge in a number of ways in order to keep the official SDK minimal, allow for future spec evolution, and support additional features. We have noted significant differences from mcp-go in the sections below. Although the API here is not compatible with mcp-go, translating between them should be straightforward in most cases. (Later, we will provide a detailed translation guide.)

Thank you to everyone who contributes to mcp-go and other Go SDKs. We hope that we can collaborate to leverage all that we've learned about MCP and Go in an official SDK.

Requirements

These may be obvious, but it's worthwhile to define goals for an official MCP SDK. An official SDK should aim to be:

• complete: it should be possible to implement every feature of the MCP spec, and these features should conform to all of the semantics described by the spec.
• idiomatic: as much as possible, MCP features should be modeled using features of the Go language and its standard library. Additionally, the SDK should repeat idioms from similar domains.
• robust: the SDK itself should be well tested and reliable, and should enable easy testability for its users.
• future-proof: the SDK should allow for future evolution of the MCP spec, in such a way that we can (as much as possible) avoid incompatible changes to the SDK API.
• extensible: to best serve the previous four concerns, the SDK should be minimal. However, it should admit extensibility using (for example) simple interfaces, middleware, or hooks.

Design considerations

In the sections below, we visit each aspect of the MCP spec, in approximately the order they are presented by the official spec For each, we discuss considerations for the Go implementation, and propose a Go API.

Foundations

Package layout

In the sections that follow, it is assumed that most of the MCP API lives in a single shared package, the mcp package. This is inconsistent with other MCP SDKs, but is consistent with Go packages like net/http, net/rpc, or google.golang.org/grpc. We believe that having a single package aids discoverability in package documentation and in the IDE. Furthermore, it avoids arbitrary decisions about package structure that may be rendered inaccurate by future evolution of the spec.

Functionality that is not directly related to MCP (like jsonschema or jsonrpc2) belongs in a separate package.

Therefore, this is the core package layout, assuming github.com/modelcontextprotocol/go-sdk as the module path.

• github.com/modelcontextprotocol/go-sdk/mcp: the bulk of the user facing API
• github.com/modelcontextprotocol/go-sdk/jsonschema: a jsonschema implementation, with validation
• github.com/modelcontextprotocol/go-sdk/internal/jsonrpc2: a fork of x/tools/internal/jsonrpc2_v2

The JSON-RPC implementation is hidden, to avoid tight coupling. As described in the next section, the only aspects of JSON-RPC that need to be exposed in the SDK are the message types, for the purposes of defining custom transports. We can expose these types by promoting them from the mcp package using aliases or wrappers.

Difference from mcp-go: Our mcp package includes all the functionality of mcp-go's mcp, client, server and transport packages.

JSON-RPC and Transports

The MCP is defined in terms of client-server communication over bidirectional JSON-RPC message streams. Specifically, version 2025-03-26 of the spec defines two transports:

• stdio: communication with a subprocess over stdin/stdout.
• streamable http: communication over a relatively complicated series of text/event-stream GET and HTTP POST requests.

Additionally, version 2024-11-05 of the spec defined a simpler (yet stateful) HTTP transport:

• sse: client issues a hanging GET request and receives messages via text/event-stream, and sends messages via POST to a session endpoint.

Furthermore, the spec states that it must be possible for users to define their own custom transports.

Given the diversity of the transport implementations, they can be challenging to abstract. However, since JSON-RPC requires a bidirectional stream, we can use this to model the MCP transport abstraction:

// A Transport is used to create a bidirectional connection between MCP client
// and server.
type Transport interface {
    Connect(ctx context.Context) (Stream, error)
}

// A Stream is a bidirectional jsonrpc2 Stream.
type Stream interface {
    Read(ctx context.Context) (JSONRPCMessage, error)
    Write(ctx context.Context, JSONRPCMessage) error
    Close() error
}

Methods accept a Go Context and return an error, as is idiomatic for APIs that do I/O.

A Transport is something that connects a logical JSON-RPC stream, and nothing more. Streams must be closeable in order to implement client and server shutdown, and therefore conform to the io.Closer interface.

Other SDKs define higher-level transports, with, for example, methods to send a notification or make a call. Those are jsonrpc2 operations on top of the logical stream, and the lower-level interface is easier to implement in most cases, which means it is easier to implement custom transports.

For our prototype, we've used an internal jsonrpc2 package based on the Go language server gopls, which we propose to fork for the MCP SDK. It already handles concerns like client/server connection, request lifecycle, cancellation, and shutdown.

Differences from mcp-go: The Go team has a battle-tested JSON-RPC implementation that we use for gopls, our Go LSP server. We are using the new version of this library as part of our MCP SDK. It handles all JSON-RPC 2.0 features, including cancellation.

The Transport interface here is lower-level than that of mcp-go, but serves a similar purpose. We believe the lower-level interface is easier to implement.

stdio transports

In the MCP Spec, the stdio transport uses newline-delimited JSON to communicate over stdin/stdout. It's possible to model both client side and server side of this communication with a shared type that communicates over an io.ReadWriteCloser. However, for the purposes of future-proofing, we should use a different types for client and server stdio transport.

The CommandTransport is the client side of the stdio transport, and connects by starting a command and binding its jsonrpc2 stream to its stdin/stdout.

// A CommandTransport is a [Transport] that runs a command and communicates
// with it over stdin/stdout, using newline-delimited JSON.
type CommandTransport struct { /* unexported fields */ }

// NewCommandTransport returns a [CommandTransport] that runs the given command
// and communicates with it over stdin/stdout.
func NewCommandTransport(cmd *exec.Command) *CommandTransport

// Connect starts the command, and connects to it over stdin/stdout.
func (*CommandTransport) Connect(ctx context.Context) (Stream, error) {

The StdIOTransport is the server side of the stdio transport, and connects by binding to os.Stdin and os.Stdout.

// A StdIOTransport is a [Transport] that communicates using newline-delimited
// JSON over stdin/stdout.
type StdIOTransport struct { /* unexported fields */ }

func NewStdIOTransport() *StdIOTransport

func (t *StdIOTransport) Connect(context.Context) (Stream, error)

HTTP transports

The HTTP transport APIs are even more asymmetrical. Since connections are initiated via HTTP requests, the client developer will create a transport, but the server developer will typically install an HTTP handler. Internally, the HTTP handler will create a logical transport for each new client connection.

Importantly, since they serve many connections, the HTTP handlers must accept a callback to get an MCP server for each new session. As described below, MCP servers can optionally connect to multiple clients. This allows customization of per-session servers: if the MCP server is stateless, the user can return the same MCP server for each connection. On the other hand, if any per-session customization is required, it is possible by returning a different Server instance for each connection.

// SSEHTTPHandler is an http.Handler that serves SSE-based MCP sessions as defined by
// the 2024-11-05 version of the MCP protocol.
type SSEHTTPHandler struct { /* unexported fields */ }

// NewSSEHTTPHandler returns a new [SSEHTTPHandler] that is ready to serve HTTP.
//
// The getServer function is used to bind created servers for new sessions. It
// is OK for getServer to return the same server multiple times.
func NewSSEHTTPHandler(getServer func(request *http.Request) *Server) *SSEHTTPHandler

func (*SSEHTTPHandler) ServeHTTP(w http.ResponseWriter, req *http.Request)

// Close prevents the SSEHTTPHandler from accepting new sessions, closes active
// sessions, and awaits their graceful termination.
func (*SSEHTTPHandler) Close() error

Notably absent are options to hook into low-level request handling for the purposes of authentication or context injection. These concerns are instead handled using standard HTTP middleware patterns. For middleware at the level of the MCP protocol, see Middleware below.

By default, the SSE handler creates messages endpoints with the ?sessionId=... query parameter. Users that want more control over the management of sessions and session endpoints may write their own handler, and create SSEServerTransport instances themselves for incoming GET requests.

// A SSEServerTransport is a logical SSE session created through a hanging GET
// request.
//
// When connected, it returns the following [Stream] implementation:
//   - Writes are SSE 'message' events to the GET response.
//   - Reads are received from POSTs to the session endpoint, via
//     [SSEServerTransport.ServeHTTP].
//   - Close terminates the hanging GET.
type SSEServerTransport struct { /* ... */ }

// NewSSEServerTransport creates a new SSE transport for the given messages
// endpoint, and hanging GET response.
//
// Use [SSEServerTransport.Connect] to initiate the flow of messages.
//
// The transport is itself an [http.Handler]. It is the caller's responsibility
// to ensure that the resulting transport serves HTTP requests on the given
// session endpoint.
//
// Most callers should instead use an [SSEHandler], which transparently handles
// the delegation to SSEServerTransports.
func NewSSEServerTransport(endpoint string, w http.ResponseWriter) *SSEServerTransport

// ServeHTTP handles POST requests to the transport endpoint.
func (*SSEServerTransport) ServeHTTP(w http.ResponseWriter, req *http.Request)

// Connect sends the 'endpoint' event to the client.
// See [SSEServerTransport] for more details on the [Stream] implementation.
func (*SSEServerTransport) Connect(context.Context) (Stream, error)

The SSE client transport is simpler, and hopefully self-explanatory.

type SSEClientTransport struct { /* ... */ }

// NewSSEClientTransport returns a new client transport that connects to the
// SSE server at the provided URL.
func NewSSEClientTransport(url string) (*SSEClientTransport, error) {

// Connect connects through the client endpoint.
func (*SSEClientTransport) Connect(ctx context.Context) (Stream, error)

The Streamable HTTP transports are similar to the SSE transport, albeit with a
more complicated implementation. For brevity, we summarize only the differences
from the equivalent SSE types:

// The StreamableHTTPHandler interface is symmetrical to the SSEHTTPHandler.
type StreamableHTTPHandler struct { /* unexported fields */ }
func NewStreamableHTTPHandler(getServer func(request *http.Request) *Server) *StreamableHTTPHandler
func (*StreamableHTTPHandler) ServeHTTP(w http.ResponseWriter, req *http.Request)
func (*StreamableHTTPHandler) Close() error

// Unlike the SSE transport, the streamable transport constructor accepts a
// session ID, not an endpoint, along with the HTTP response for the request
// that created the session. It is the caller's responsibility to delegate
// requests to this session.
type StreamableServerTransport struct { /* ... */ }
func NewStreamableServerTransport(sessionID string, w http.ResponseWriter) *StreamableServerTransport
func (*StreamableServerTransport) ServeHTTP(w http.ResponseWriter, req *http.Request)
func (*StreamableServerTransport) Connect(context.Context) (Stream, error)

// The streamable client handles reconnection transparently to the user.
type StreamableClientTransport struct { /* ... */ }
func NewStreamableClientTransport(url string) *StreamableClientTransport {
func (*StreamableClientTransport) Connect(context.Context) (Stream, error)

Differences from mcp-go: In mcp-go, server authors create an MCPServer, populate it with tools, resources and so on, and then wrap it in an SSEServer or StdioServer. Users can manage their own sessions with RegisterSession and UnregisterSession. Rather than use a server constructor to get a distinct server for each connection, there is a concept of a "session tool" that overlays tools for a specific session.

Here, we tried to differentiate the concept of a Server, HTTPHandler, and Transport, and provide per-session customization through either the getServer constructor or middleware. Additionally, individual handlers and transports here have a minimal API, and do not expose internal details. (Open question: are we oversimplifying?)

Other transports

We also provide a couple of transport implementations for special scenarios. An InMemoryTransport can be used when the client and server reside in the same process. A LoggingTransport is a middleware layer that logs RPC logs to a desired location, specified as an io.Writer.

// An InMemoryTransport is a [Transport] that communicates over an in-memory
// network connection, using newline-delimited JSON.
type InMemoryTransport struct { /* ... */ }

// NewInMemoryTransports returns two InMemoryTransports that connect to each
// other.
func NewInMemoryTransports() (*InMemoryTransport, *InMemoryTransport)

// A LoggingTransport is a [Transport] that delegates to another transport,
// writing RPC logs to an io.Writer.
type LoggingTransport struct { /* ... */ }
func NewLoggingTransport(delegate Transport, w io.Writer) *LoggingTransport

Protocol types

Types needed for the protocol are generated from the JSON schema of the MCP spec.

These types will be included in the mcp package, but will be unexported unless they are needed for the user-facing API. Notably, JSON-RPC request types are elided, since they are handled by the jsonrpc2 package and should not be observed by the user.

For user-provided data, we use json.RawMessage, so that marshalling/unmarshalling can be delegated to the business logic of the client or server.

For union types, which can't be represented in Go (specifically Content and ResourceContents), we prefer distinguished unions: struct types with fields corresponding to the union of all properties for union elements.

For brevity, only a few examples are shown here:

type ReadResourceParams struct {
	URI string `json:"uri"`
}

type CallToolResult struct {
	Meta    map[string]json.RawMessage `json:"_meta,omitempty"`
	Content []Content                  `json:"content"`
	IsError bool                       `json:"isError,omitempty"`
}

// Content is the wire format for content.
//
// The Type field distinguishes the type of the content.
// At most one of Text, MIMEType, Data, and Resource is non-zero.
type Content struct {
	Type     string            `json:"type"`
	Text     string            `json:"text,omitempty"`
	MIMEType string            `json:"mimeType,omitempty"`
	Data     []byte            `json:"data,omitempty"`
	Resource *ResourceContents `json:"resource,omitempty"`
}

// NewTextContent creates a [Content] with text.
func NewTextContent(text string) *Content
// etc.

Differences from mcp-go: these types are largely similar, but our type generator flattens types rather than using struct embedding.

Clients and Servers

Generally speaking, the SDK is used by creating a Client or Server instance, adding features to it, and connecting it to a peer.

However, the SDK must make a non-obvious choice in these APIs: are clients 1:1 with their logical connections? What about servers? Both clients and servers are stateful: users may add or remove roots from clients, and tools, prompts, and resources from servers. Additionally, handlers for these features may themselves be stateful, for example if a tool handler caches state from earlier requests in the session.

We believe that in the common case, any change to a client or server, such as adding a tool, is intended for all its peers. It is therefore more useful to allow multiple connections from a client, and to a server. This is similar to the net/http packages, in which an http.Client and http.Server each may handle multiple unrelated connections. When users add features to a client or server, all connected peers are notified of the change.

Supporting multiple connections to servers (and from clients) still allows for stateful components, as it is up to the user to decide whether or not to create distinct servers/clients for each connection. For example, if the user wants to create a distinct server for each new connection, they can do so in the getServer factory passed to transport handlers.

Following the terminology of the spec, we call the logical connection between a client and server a "session." There must necessarily be a ClientSession and a ServerSession, corresponding to the APIs available from the client and server perspective, respectively.

Client                                                   Server
  ⇅                          (jsonrpc2)                     ⇅
ClientSession ⇄ Client Transport ⇄ Server Transport ⇄ ServerSession

Sessions are created from either Client or Server using the Connect method.

type Client struct { /* ... */ }
func NewClient(name, version string, opts *ClientOptions) *Client
func (*Client) Connect(context.Context, Transport) (*ClientSession, error)
func (*Client) Sessions() iter.Seq[*ClientSession]
// Methods for adding/removing client features are described below.

type ClientOptions struct { /* ... */ } // described below

type ClientSession struct { /* ... */ }
func (*ClientSession) Client() *Client
func (*ClientSession) Close() error
func (*ClientSession) Wait() error
// Methods for calling through the ClientSession are described below.
// For example: ClientSession.ListTools.

type Server struct { /* ... */ }
func NewServer(name, version string, opts *ServerOptions) *Server
func (*Server) Connect(context.Context, Transport) (*ServerSession, error)
func (*Server) Sessions() iter.Seq[*ServerSession]
// Methods for adding/removing server features are described below.

type ServerOptions struct { /* ... */ } // described below

type ServerSession struct { /* ... */ }
func (*ServerSession) Server() *Server
func (*ServerSession) Close() error
func (*ServerSession) Wait() error
// Methods for calling through the ServerSession are described below.
// For example: ServerSession.ListRoots.

Here's an example of these APIs from the client side:

client := mcp.NewClient("mcp-client", "v1.0.0", nil)
// Connect to a server over stdin/stdout
transport := mcp.NewCommandTransport(exec.Command("myserver"))
session, err := client.Connect(ctx, transport)
if err != nil { ... }
// Call a tool on the server.
content, err := session.CallTool(ctx, "greet", map[string]any{"name": "you"}, nil)
...
return session.Close()

A server that can handle that client call would look like this:

// Create a server with a single tool.
server := mcp.NewServer("greeter", "v1.0.0", nil)
server.AddTools(mcp.NewTool("greet", "say hi", SayHi))
// Run the server over stdin/stdout, until the client disconnects.
transport := mcp.NewStdIOTransport()
session, err := server.Connect(ctx, transport)
...
return session.Wait()

For convenience, we provide Server.Run to handle the common case of running a session until the client disconnects:

func (*Server) Run(context.Context, Transport)

Differences from mcp-go: the Server APIs are similar to mcp-go, though the association between servers and transports is different. In mcp-go, a single server is bound to what we would call an SSEHTTPHandler, and reused for all sessions. Per-session behavior is implemented though a 'session tool' overlay. As discussed above, the transport abstraction here is differentiated from HTTP serving, and the Server.Connect method provides a consistent API for binding to an arbitrary transport. Servers here do not have methods for sending notifications or calls, because they are logically distinct from the ServerSession. In mcp-go, servers are n:1, but there is no abstraction of a server session: sessions are addressed in Server APIs through their sessionID: SendNotificationToAllClients, SendNotificationToClient, SendNotificationToSpecificClient.

The client API here is different, since clients and client sessions are conceptually distinct. The ClientSession is closer to mcp-go's notion of Client.

For both clients and servers, mcp-go uses variadic options to customize behavior, whereas an options struct is used here. We felt that in this case, an options struct would be more readable, and result in simpler package documentation.

Spec Methods

In our SDK, RPC methods that are defined in the specification take a context and a params pointer as arguments, and return a result pointer and error:

func (*ClientSession) ListTools(context.Context, *ListToolsParams) (*ListToolsResult, error)

Our SDK has a method for every RPC in the spec, and except for CallTool, their signatures all share this form. We do this, rather than providing more convenient shortcut signatures, to maintain backward compatibility if the spec makes backward-compatible changes such as adding a new property to the request parameters (as in this commit, for example). To avoid boilerplate, we don't repeat this signature for RPCs defined in the spec; readers may assume it when we mention a "spec method."

CallTool is the only exception: for convenience, it takes the tool name and arguments, with an options struct for additional request fields. See the section on Tools below for details.

Why do we use params instead of the full JSON-RPC request? As much as possible, we endeavor to hide JSON-RPC details when they are not relevant to the business logic of your client or server. In this case, the additional information in the JSON-RPC request is just the request ID and method name; the request ID is irrelevant, and the method name is implied by the name of the Go method providing the API.

We believe that any change to the spec that would require callers to pass a new a parameter is not backward compatible. Therefore, it will always work to pass nil for any XXXParams argument that isn't currently necessary. For example, it is okay to call Ping like so:

err := session.Ping(ctx, nil)

Iterator Methods

For convenience, iterator methods handle pagination for the List spec methods automatically, traversing all pages. If Params are supplied, iteration begins from the provided cursor (if present).

func (*ClientSession) Tools(context.Context, *ListToolsParams) iter.Seq2[Tool, error]

func (*ClientSession) Prompts(context.Context, *ListPromptsParams) iter.Seq2[Prompt, error]

func (*ClientSession) Resources(context.Context, *ListResourceParams) iter.Seq2[Resource, error]

func (*ClientSession) ResourceTemplates(context.Context, *ListResourceTemplatesParams) iter.Seq2[ResourceTemplate, error]

Middleware

We provide a mechanism to add MCP-level middleware on the both the client and server side, which runs after the request has been parsed but before any normal handling.

// A MethodHandler handles MCP messages.
// The params argument is an XXXParams struct pointer, such as *GetPromptParams.
// For methods, a MethodHandler must return either an XXResult struct pointer and a nil error, or
// nil with a non-nil error.
// For notifications, a MethodHandler must return nil, nil.
type MethodHandler[S ClientSession | ServerSession] func(
	ctx context.Context, _ *S, method string, params any) (result any, err error)

// Middleware is a function from MethodHandlers to MethodHandlers.
type Middleware[S ClientSession | ServerSession] func(MethodHandler[S]) MethodHandler[S]

// AddMiddleware wraps the client/server's current method handler using the provided
// middleware. Middleware is applied from right to left, so that the first one
// is executed first.
//
// For example, AddMiddleware(m1, m2, m3) augments the server method handler as
// m1(m2(m3(handler))).
func (c *Client) AddMiddleware(middleware ...Middleware[ClientSession])
func (s *Server) AddMiddleware(middleware ...Middleware[ServerSession])

As an example, this code adds server-side logging:

func withLogging(h mcp.MethodHandler[ServerSession]) mcp.MethodHandler[ServerSession]{
    return func(ctx context.Context, s *mcp.ServerSession, method string, params any) (res any, err error) {
        log.Printf("request: %s %v", method, params)
        defer func() { log.Printf("response: %v, %v", res, err) }()
        return h(ctx, s , method, params)
    }
}

server.AddMiddleware(withLogging)

Differences from mcp-go: Version 0.26.0 of mcp-go defines 24 server hooks. Each hook consists of a field in the Hooks struct, a Hooks.Add method, and a type for the hook function. These are rarely used. The most common is OnError, which occurs fewer than ten times in open-source code.

Errors

With the exception of tool handler errors, protocol errors are handled transparently as Go errors: errors in server-side feature handlers are propagated as errors from calls from the ClientSession, and vice-versa.

Protocol errors wrap a JSONRPCError type which exposes its underlying error code.

type JSONRPCError struct {
	Code    int64           `json:"code"`
	Message string          `json:"message"`
	Data    json.RawMessage `json:"data,omitempty"`
}

As described by the spec, tool execution errors are reported in tool results.

Differences from mcp-go: the JSONRPCError type here does not include ID and Method, which can be inferred from the caller. Otherwise, this behavior is similar.

Cancellation

Cancellation is implemented transparently using context cancellation. The user can cancel an operation by cancelling the associated context:

ctx, cancel := context.WithCancel(ctx)
go session.CallTool(ctx, "slow", map[string]any{}, nil)
cancel()

When this client call is cancelled, a "notifications/cancelled" notification is sent to the server. However, the client call returns immediately with ctx.Err(): it does not wait for the result from the server.

The server observes a client cancellation as a cancelled context.

Progress handling

A caller can request progress notifications by setting the ProgressToken field on any request.

type XXXParams struct { // where XXX is each type of call
  ...
  ProgressToken any // string or int
}

Handlers can notify their peer about progress by calling the NotifyProgress method. The notification is only sent if the peer requested it by providing a progress token.

func (*ClientSession) NotifyProgress(context.Context, *ProgressNotification)
func (*ServerSession) NotifyProgress(context.Context, *ProgressNotification)

Ping / KeepAlive

Both ClientSession and ServerSession expose a Ping method to call "ping" on their peer.

func (c *ClientSession) Ping(ctx context.Context, *PingParams) error
func (c *ServerSession) Ping(ctx context.Context, *PingParams) error

Additionally, client and server sessions can be configured with automatic keepalive behavior. If the KeepAlive option is set to a non-zero duration, it defines an interval for regular "ping" requests. If the peer fails to respond to pings originating from the keepalive check, the session is automatically closed.

type ClientOptions struct {
  ...
  KeepAlive time.Duration
}

type ServerOptions struct {
  ...
  KeepAlive time.Duration
}

Differences from mcp-go: in mcp-go the Ping method is only provided for client, not server, and the keepalive option is only provided for SSE servers (as a variadic option).

Client Features

Roots

Clients support the MCP Roots feature, including roots-changed notifications. Roots can be added and removed from a Client with AddRoots and RemoveRoots:

// AddRoots adds the given roots to the client,
// replacing any with the same URIs,
// and notifies any connected servers.
func (*Client) AddRoots(roots ...*Root)

// RemoveRoots removes the roots with the given URIs.
// and notifies any connected servers if the list has changed.
// It is not an error to remove a nonexistent root.
func (*Client) RemoveRoots(uris ...string)

Server sessions can call the spec method ListRoots to get the roots. If a server installs a RootsChangedHandler, it will be called when the client sends a roots-changed notification, which happens whenever the list of roots changes after a connection has been established.

type ServerOptions {
  ...
  // If non-nil, called when a client sends a roots-changed notification.
  RootsChangedHandler func(context.Context, *ServerSession, *RootsChangedParams)
}

The Roots method provides a cached iterator of the root set, invalidated when roots change.

func (*ServerSession) Roots(context.Context) (iter.Seq[*Root, error])

Sampling

Clients that support sampling are created with a CreateMessageHandler option for handling server calls. To perform sampling, a server session calls the spec method CreateMessage.

type ClientOptions struct {
  ...
  CreateMessageHandler func(context.Context, *ClientSession, *CreateMessageParams) (*CreateMessageResult, error)
}

Server Features

Tools

A Tool is a logical MCP tool, generated from the MCP spec, and a ServerTool is a tool bound to a tool handler.

type Tool struct {
	Annotations *ToolAnnotations   `json:"annotations,omitempty"`
	Description string             `json:"description,omitempty"`
	InputSchema *jsonschema.Schema `json:"inputSchema"`
	Name string                    `json:"name"`
}

type ToolHandler func(context.Context, *ServerSession, *CallToolParams) (*CallToolResult, error)

type ServerTool struct {
	Tool    Tool
	Handler ToolHandler
}

Add tools to a server with AddTools:

server.AddTools(
  mcp.NewTool("add", "add numbers", addHandler),
  mcp.NewTool("subtract, subtract numbers", subHandler))

Remove them by name with RemoveTools:

server.RemoveTools("add", "subtract")

A tool's input schema, expressed as a JSON Schema, provides a way to validate the tool's input. One of the challenges in defining tools is the need to associate them with a Go function, yet support the arbitrary complexity of JSON Schema. To achieve this, we have seen two primary approaches:

1. Use reflection to generate the tool's input schema from a Go type (à la metoro-io/mcp-golang)
2. Explicitly build the input schema (à la mark3labs/mcp-go).

Both of these have their advantages and disadvantages. Reflection is nice, because it allows you to bind directly to a Go API, and means that the JSON schema of your API is compatible with your Go types by construction. It also means that concerns like parsing and validation can be handled automatically. However, it can become cumbersome to express the full breadth of JSON schema using Go types or struct tags, and sometimes you want to express things that aren’t naturally modeled by Go types, like unions. Explicit schemas are simple and readable, and give the caller full control over their tool definition, but involve significant boilerplate.

We have found that a hybrid model works well, where the initial schema is derived using reflection, but any customization on top of that schema is applied using variadic options. We achieve this using a NewTool helper, which generates the schema from the input type, and wraps the handler to provide parsing and validation. The schema (and potentially other features) can be customized using ToolOptions.

// NewTool creates a Tool using reflection on the given handler.
func NewTool[TInput any](name, description string, handler func(context.Context, *ServerSession, TInput) ([]Content, error), opts …ToolOption) *ServerTool

type ToolOption interface { /* ... */ }

NewTool determines the input schema for a Tool from the struct used in the handler. Each struct field that would be marshaled by encoding/json.Marshal becomes a property of the schema. The property is required unless the field's json tag specifies "omitempty" or "omitzero" (new in Go 1.24). For example, given this struct:

struct {
  Name     string `json:"name"`
  Count    int    `json:"count,omitempty"`
  Choices  []string
  Password []byte `json:"-"`
}

"name" and "Choices" are required, while "count" is optional.

As of this writing, the only ToolOption is Input, which allows customizing the input schema of the tool using schema options. These schema options are recursive, in the sense that they may also be applied to properties.

func Input(...SchemaOption) ToolOption

type Property(name string, opts ...SchemaOption) SchemaOption
type Description(desc string) SchemaOption
// etc.

For example:

NewTool(name, description, handler,
    Input(Property("count", Description("size of the inventory"))))

The most recent JSON Schema spec defines over 40 keywords. Providing them all as options would bloat the API despite the fact that most would be very rarely used. For less common keywords, use the Schema option to set the schema explicitly:

NewTool(name, description, handler,
    Input(Property("Choices", Schema(&jsonschema.Schema{UniqueItems: true}))))

Schemas are validated on the server before the tool handler is called.

Since all the fields of the Tool struct are exported, a Tool can also be created directly with assignment or a struct literal.

Client sessions can call the spec method ListTools or an iterator method Tools to list the available tools.

As mentioned above, the client session method CallTool has a non-standard signature, so that CallTool can handle the marshalling of tool arguments: the type of CallToolParams.Arguments is json.RawMessage, to delegate unmarshalling to the tool handler.

func (c *ClientSession) CallTool(ctx context.Context, name string, args map[string]any, opts *CallToolOptions) (_ *CallToolResult, err error)

type CallToolOptions struct {
	ProgressToken any // string or int
}

Differences from mcp-go: using variadic options to configure tools was significantly inspired by mcp-go. However, the distinction between ToolOption and SchemaOption allows for recursive application of schema options. For example, that limitation is visible in this code, which must resort to untyped maps to express a nested schema.

Additionally, the NewTool helper provides a means for building a tool from a Go function using reflection, that automatically handles parsing and validation of inputs.

We provide a full JSON Schema implementation for validating tool input schemas against incoming arguments. The jsonschema.Schema type provides exported features for all keywords in the JSON Schema draft2020-12 spec. Tool definers can use it to construct any schema they want, so there is no need to provide options for all of them. When combined with schema inference from input structs, we found that we needed only three options to cover the common cases, instead of mcp-go's 23. For example, we will provide Enum, which occurs 125 times in open source code, but not MinItems, MinLength or MinProperties, which each occur only once (and in an SDK that wraps mcp-go).

For registering tools, we provide only AddTools; mcp-go's SetTools, AddTool, AddSessionTool, and AddSessionTools are deemed unnecessary. (Similarly for Delete/Remove).

Prompts

Use NewPrompt to create a prompt. As with tools, prompt argument schemas can be inferred from a struct, or obtained from options.

func NewPrompt[TReq any](name, description string,
  handler func(context.Context, *ServerSession, TReq) (*GetPromptResult, error),
  opts ...PromptOption) *ServerPrompt

Use AddPrompts to add prompts to the server, and RemovePrompts
to remove them by name.

type codeReviewArgs struct {
  Code string `json:"code"`
}

func codeReviewHandler(context.Context, *ServerSession, codeReviewArgs) {...}

server.AddPrompts(
  NewPrompt("code_review", "review code", codeReviewHandler,
    Argument("code", Description("the code to review"))))

server.RemovePrompts("code_review")

Client sessions can call the spec method ListPrompts or the iterator method Prompts to list the available prompts, and the spec method GetPrompt to get one.

Differences from mcp-go: We provide a NewPrompt helper to bind a prompt handler to a Go function using reflection to derive its arguments. We provide RemovePrompts to remove prompts from the server.

Resources and resource templates

In our design, each resource and resource template is associated with a function that reads it, with this signature:

type ResourceHandler func(context.Context, *ServerSession, *ReadResourceParams) (*ReadResourceResult, error)

The arguments include the ServerSession so the handler can observe the client's roots. The handler should return the resource contents in a ReadResourceResult, calling either NewTextResourceContents or NewBlobResourceContents. If the handler omits the URI or MIME type, the server will populate them from the resource.

The ServerResource and ServerResourceTemplate types hold the association between the resource and its handler:

type ServerResource struct {
  Resource Resource
  Handler  ResourceHandler
}

type ServerResourceTemplate struct {
  Template ResourceTemplate
  Handler  ResourceHandler
}

To add a resource or resource template to a server, users call the AddResources and AddResourceTemplates methods with one or more ServerResources or ServerResourceTemplates. We also provide methods to remove them.

func (*Server) AddResources(...*ServerResource)
func (*Server) AddResourceTemplates(...*ServerResourceTemplate)

func (s *Server) RemoveResources(uris ...string)
func (s *Server) RemoveResourceTemplates(uriTemplates ...string)

The ReadResource method finds a resource or resource template matching the argument URI and calls its associated handler.

To read files from the local filesystem, we recommend using FileResourceHandler to construct a handler:

// FileResourceHandler returns a ResourceHandler that reads paths using dir as a root directory.
// It protects against path traversal attacks.
// It will not read any file that is not in the root set of the client requesting the resource.
func (*Server) FileResourceHandler(dir string) ResourceHandler

Here is an example:

// Safely read "/public/puppies.txt".
s.AddResources(&mcp.ServerResource{
  Resource: mcp.Resource{URI: "file:///puppies.txt"},
  Handler: s.FileReadResourceHandler("/public")})

Server sessions also support the spec methods ListResources and ListResourceTemplates, and the corresponding iterator methods Resources and ResourceTemplates.

Differences from mcp-go: for symmetry with tools and prompts, we use AddResources rather than AddResource. Additionally, the ResourceHandler returns a ReadResourceResult, rather than just its content, for compatibility with future evolution of the spec.

Subscriptions

ClientSessions can manage change notifications on particular resources:

func (*ClientSession) Subscribe(context.Context, *SubscribeParams) error
func (*ClientSession) Unsubscribe(context.Context, *UnsubscribeParams) error

The server does not implement resource subscriptions. It passes along subscription requests to the user, and supplies a method to notify clients of changes. It tracks which sessions have subscribed to which resources so the user doesn't have to.

If a server author wants to support resource subscriptions, they must provide handlers to be called when clients subscribe and unsubscribe. It is an error to provide only one of these handlers.

type ServerOptions struct {
  ...
  // Function called when a client session subscribes to a resource.
  SubscribeHandler func(context.Context, *SubscribeParams) error
  // Function called when a client session unsubscribes from a resource.
  UnsubscribeHandler func(context.Context, *UnsubscribeParams) error
}

User code should call ResourceUpdated when a subscribed resource changes.

func (*Server) ResourceUpdated(context.Context, *ResourceUpdatedNotification) error

The server routes these notifications to the server sessions that subscribed to the resource.

ListChanged notifications

When a list of tools, prompts or resources changes as the result of an AddXXX or RemoveXXX call, the server informs all its connected clients by sending the corresponding type of notification. A client will receive these notifications if it was created with the corresponding option:

type ClientOptions struct {
  ...
  ToolListChangedHandler func(context.Context, *ClientSession, *ToolListChangedParams)
  PromptListChangedHandler func(context.Context, *ClientSession, *PromptListChangedParams)
  // For both resources and resource templates.
  ResourceListChangedHandler func(context.Context, *ClientSession, *ResourceListChangedParams)
}

Differences from mcp-go: mcp-go instead provides a general OnNotification handler. For type-safety, and to hide JSON RPC details, we provide feature-specific handlers here.

Completion

Clients call the spec method Complete to request completions. Servers automatically handle these requests based on their collections of prompts and resources.

Differences from mcp-go: the client API is similar. mcp-go has not yet defined its server-side behavior.

Logging

MCP specifies a notification for servers to log to clients. Server sessions implement this with the LoggingMessage method. It honors the minimum log level established by the client session's SetLevel call.

As a convenience, we also provide a slog.Handler that allows server authors to write logs with the log/slog package::

// A LoggingHandler is a [slog.Handler] for MCP.
type LoggingHandler struct {...}

// LoggingHandlerOptions are options for a LoggingHandler.
type LoggingHandlerOptions struct {
	// The value for the "logger" field of logging notifications.
	LoggerName string
	// Limits the rate at which log messages are sent.
	// If zero, there is no rate limiting.
	MinInterval time.Duration
}

// NewLoggingHandler creates a [LoggingHandler] that logs to the given [ServerSession] using a
// [slog.JSONHandler].
func NewLoggingHandler(ss *ServerSession, opts *LoggingHandlerOptions) *LoggingHandler

Server-to-client logging is configured with ServerOptions:

type ServerOptions {
  ...
  // The value for the "logger" field of the notification.
  LoggerName string
  // Log notifications to a single ClientSession will not be
  // sent more frequently than this duration.
  LoggingInterval time.Duration
}

A call to a log method like Info is translated to a LoggingMessageNotification as follows:

• The attributes and the message populate the "data" property with the output of a slog.JSONHandler: The result is always a JSON object, with the key "msg" for the message.

• If the LoggerName server option is set, it populates the "logger" property.

• The standard slog levels Info, Debug, Warn and Error map to the corresponding levels in the MCP spec. The other spec levels map to integers between the slog levels. For example, "notice" is level 2 because it is between "warning" (slog value 4) and "info" (slog value 0). The mcp package defines consts for these levels. To log at the "notice" level, a handler would call Log(ctx, mcp.LevelNotice, "message").

A client that wishes to receive log messages must provide a handler:

type ClientOptions struct {
  ...
  LoggingMessageHandler func(context.Context, *ClientSession, *LoggingMessageParams)
}

Pagination

Servers initiate pagination for ListTools, ListPrompts, ListResources, and ListResourceTemplates, dictating the page size and providing a NextCursor field in the Result if more pages exist. The SDK implements keyset pagination, using the unique ID of the feature as the key for a stable sort order and encoding the cursor as an opaque string.

For server implementations, the page size for the list operation may be configured via the ServerOptions.PageSize field. PageSize must be a non-negative integer. If zero, a sensible default is used.

type ServerOptions {
  ...
  PageSize int
}

Client requests for List methods include an optional Cursor field for pagination. Server responses for List methods include a NextCursor field if more pages exist.

In addition to the List methods, the SDK provides an iterator method for each list operation. This simplifies pagination for clients by automatically handling the underlying pagination logic. See Iterator Methods above.

Differences with mcp-go: the PageSize configuration is set with a configuration field rather than a variadic option. Additionally, this design proposes pagination by default, as this is likely desirable for most servers

Scope

• Protocol Specification
• SDK Features
• Documentation
• Developer Experience
• Other

[SamMorrowDrums]

Thank you so much for hard work on this!

Excited to try this out at GitHub. 🔥

[williammartin]

This looks great, thank you for your work. As I read the function signatures, I had a similar question as: mark3labs/mcp-go#294, which perhaps you also have a good answer for.

You say:

As described by the spec, tool execution errors are reported in tool results.

And the ToolHandler has signature:

type ToolHandler func(context.Context, *ServerSession, *CallToolParams) (*CallToolResult, error)

Under what circumstances would you expect a Tool to return a non-nil error as the second return value? Thanks!

[findleyr]

@williammartin here's the relevant implementation in our prototype:
https://go.googlesource.com/tools/+/refs/heads/master/internal/mcp/tool.go#37

If you use the NewTool helper (which binds a tool to a Go function), errors in that function are lifted into the CallToolResult, as prescribed by the spec. If you write your own ToolHandler, you'd have to implement the equivalent logic yourself. Let me know if you think that's not correct.

[MegaGrindStone]

Some concrete example of returned error would be, in the filesystem mcp server, if your handler unable to read or write the file requested by the client, you would returns the error and explained what's happen to the client.

But, I want to clarify something to @findleyr. Based on the implementation you share above, if the handler implementation is returning error, the client that calling it would NOT receive a Go error, but instead a CallToolResult with an IsError field set to true, right?

[findleyr]

@MegaGrindStone yes, that's right. That was my reading of the spec: a missing tool or parse error results in a JSON-RPC error, but a failure of the business logic of the tool (the Go function) is lifted into a normal result with "isError": true.

[williammartin]

Thanks @findleyr for sharing that code! That's pretty much exactly what I imagined earlier in a conversation when I wrote to @SamMorrowDrums about separating our tool handling from the specific server module:

So for example, we might have our own handler func(context.Context, args map[string]any) ToolCallResult (glossing over some unknowns about what handlers actually need from the request, maybe we just have our own request type) where ToolCallResult is a sealed interface. Then we can swap out the MCP implementation very easily by just adjusting the mapping layer

Except in this implementation you've chosen to union fields rather than types (totally fine, I'm just a sum typer at heart!). Also, a big fan of giving access to the raw json for input, thats what I would have used in the signature above if it wouldn't have meant marshaling the mcp-go args again.

[williammartin]

@findleyr, as valuable as this document is for noodling on, I'm very eager to get my hands on the implementation to give feedback based on empirical use. At the top of the discussion you said:

We hope that we can reach consensus within a reasonable amount of time (a couple of weeks, perhaps?). Once that happens, we’ll finalize the design and move to the execution phase, where we create github.com/modelcontextprotocol/go-sdk.

Please let me know if you start executing in such a way that I can start trying it out.

[RazaGR]

Really excited to see this moving forward — the design looks thoughtful and well aligned with Go idioms. I’d be happy to help with early testing, especially around session handling, transport abstractions, and tool integration. Please feel free to loop me in where hands-on feedback would be most useful. Looking forward to contributing!

[anuraaga]

Thanks for the doc @findleyr. As you know I am mostly looking at the observability aspect of the SDK, especially propagating distributed context. Some initial thoughts

• As discussed in https://github.com/orgs/modelcontextprotocol/discussions/224#discussioncomment-13095826, we will need a way to handle _meta generically. You mentioned needing to customize generation logic and I was wondering if it's possible to add some details about it to the doc. Also some notable questions are

  • How would it be accessed, i.e. in middleware? Would we need a huge type switch? Or perhaps it makes sense to make params a generic type, e.g. Params[CallToolParams], func (p *Params) Meta(), etc
  • Or maybe meta can be recognized for what it is and passed within Context and retrieved something like mcp.MetaFromContext(ctx). This seems particularly important if the proposal to move meta to the top level goes through Document request.params._meta convention modelcontextprotocol#414 (comment) as it seems we wouldn't want to have to add meta to every single function that accepts params

• A bit related to the above, there seems to be a notion that all types in MCP are actually open meaning they should provide access to additional fields. Does it make sense to make every type define custom marshaling to store unknown fields into a map? Of course I wish Go had a json: struct tag to store them more easily

  • I raised feat(protocol): allow additional fields in meta mark3labs/mcp-go#293 to mcp-go specifically for _meta since it is the one that has current use as an open container and I didn't want to do too intrusive of a change. But if you follow the links to the other SDKs from there, you'll see that they allow extension on almost all types, not just _meta

• Middleware seems to only be defined for server. However, MCP supports making requests from the server to the client. An example here. So to make sure the client handling such a request can receive context from the server, I think middleware needs to be addable in a symmetric way between client/server. In my instrumentation of SDKs I have tried to instrument Transport for this reason - it worked well in Python and Typescript because the way async iterators work, they actually handle the messages they receive and can setup context. It didn't work for Java, though I made a suggestion to refactor in a way that potentially could. If wanting to make Transport a symmetric entrypoint for wrapping with context, I think in Go it may look like a Write() and ReadLoop(), the latter not actually returning a message but looping on received messages to pass to a JSON-RPC handler that could be instrumented with middleware. If Transport isn't the place to think about this though, hopefully there is a different way to do middleware symmetric to server and client.

  • BTW, to clarify, if we ignore server -> client requests for now, I guess the intention of the design is to use server middleware to read context, and wrap Transport to wrap Stream on the client to write it?

[jba]

Here's the CL (PR) for middleware: https://go-review.git.corp.google.com/c/tools/+/673515.
Basically matches the design doc, with some changed names.
We will do client-side middleware too.

We are working on a design for meta, but it will definitely be open. (There will be a map[string]any somewhere.) You make a good point about being able to access Meta without known the concrete type. We'll consider that.

We also have to talk about making all types open. My gut feeling is no: the spec is clear about what's open and not, and making a closed type open invites incompatibilites over the wire.

[anuraaga]

Thanks - I read through the latest code and see the symmetric middleware and the request handling side looks pretty good. I am still curious what it'll look like to inject data into messages before going over the wire, I think it's wrap Transport to wrap Stream, but some concerns are that it is a bit tedious to wrap two types - but that's fine really. More worryingly is if these interfaces may be extended using optional interfaces in the future like http.ResponseWriter. Anyone that has written net/http middleware will know how incredibly difficult it is to wrap that type, while being very common to do - if the Transport interface seems pretty stable and we optimistically expect no optional interfaces, even when it happens it could be ok but want to call that out as a concern.

Also, for requests from server to client, or arbitrary request from client/server, I think it can be simply added with AddRequestHandler and Session.SendRequest or similar. I don't know if such functionality would be needed for an initial release but checking on whether this sort of design is on the table for future work to support it.

[anuraaga]

I went ahead and prototyped what it would look like to inject metadata by wrapping Transport. I realized one issue was that the params are already serialized into json.RawMessage by the time they hit the transport, so injecting means unmarshalling, injecting, and remarshalling. It's doable but I guess it's nice if this could happen before the original marshalling somehow.

That being said, since it is doable, once meta is available on the server side things look ready for context propagation to work so nice to see that.

type contextTransport struct {
	delegate mcp.Transport
}

func (c contextTransport) Connect(ctx context.Context) (mcp.Stream, error) {
	stream, err := c.delegate.Connect(ctx)
	if err != nil {
		return nil, err
	}
	return contextStream{delegate: stream}, nil
}

type contextStream struct {
	delegate mcp.Stream
}

// Read implements mcp.Stream.
func (c contextStream) Read(ctx context.Context) (jsonrpc2.Message, int64, error) {
	return c.delegate.Read(ctx)
}

// Write implements mcp.Stream.
func (c contextStream) Write(ctx context.Context, msg jsonrpc2.Message) (int64, error) {
	if req, ok := msg.(*jsonrpc2.Request); ok {
		var params map[string]any
		if err := json.Unmarshal(req.Params, &params); err == nil {
			meta := params["_meta"]
			if meta == nil {
				meta = make(map[string]any)
				params["_meta"] = meta
			}
			otel.GetTextMapPropagator().Inject(ctx, stringAnyMapCarrier{meta.(map[string]any)})
			if newParams, err := json.Marshal(params); err == nil {
				req.Params = newParams
			}
		}
	}
	return c.delegate.Write(ctx, msg)
}

// Close implements mcp.Stream.
func (c contextStream) Close() error {
	return c.delegate.Close()
}

[jba]

Ah, I see: our middleware intercepts on the receiving side (like all HTTP middleware). You need to intercept on the sending side. There are other use cases for that too, like adding a progress token to every outgoing request.
I think we need another kind of middleware, before each call, with unmarshalled params.
I'll think about that.

[tarunanand-dev]

Hello @findleyr and fellow contributors,

Thanks for the very thoughtful design proposal and discussion.

I have an alternate proposal and discussion that I kickstarted with the rust sdk authors here
https://github.com/orgs/modelcontextprotocol/discussions/354

Reproducing the gist of that discussion here:

Folks, as MCP’s language ecosystem expands, we’re spending more time re‑implementing the same deterministic‑state logic than shipping new ideas. I’d like us to flip that ratio by standardizing on a single, Rust core (rust‑sdk) that owns the workflow engine— sessions, peering, transport, serialization—while each language surface (Python, JS/TS, Go, etc.) stays 100 % idiomatic and lightweight. 

Rust gives us memory‑safe, no‑GC performance overhead, a stable FFI, and a proven path. The payoff is immediate: one quality, security and performance audit instead of five, feature parity across languages on day 1, and p99 latencies that drop from milliseconds to microseconds—without asking Python or JavaScript developers to learn Rust. 

In short, one deterministic engine, many friendly faces. We have built a V 0.1 Python and Typescript SDK based on this architecture so that we write less boilerplate and innovate faster on the parts that matter.

I know that as Golang stewards and enthusiasts you may want to write the SDK from scratch but would love to hear your views and counterpoints to this.

The Python Bindings PR is here
modelcontextprotocol/rust-sdk#172

The Typescript Bindings PR is here
modelcontextprotocol/rust-sdk#183

[anuraaga]

Thanks @tarunanand-dev - just one piece of advice is to be clear that the tradeoffs aren't really technical but between user experience vs maintainer experience, because a native core has a tendency to reduce the user experience at least to some degree. It's a discussion that comes up in any OSS community but I have seen the ones that are most successful are the ones that double down on user experience - the linked thread actually brings one up, grpc which is great tech but generally not great user experience and I think keeps it to a very limited audience (like me, but I only use connectrpc nowadays which is a very similar and also compatible protocol in user friendly sdks).

I remember many problems with grpc for Node before they migrated from a C++ core to pure-JS and it made it usable. grpc-python continues to be native and when it has problems it can be really hard to debug. Users do debug their apps ;)

Don't want to discourage the idea of a shared native core! And I do think grpc is only a good reference point and not equivalent since it's pretty old and likely has poorly designed FFI. But I do want to make sure there is enough consideration on the implication on user experience, not just technical issues. Personally, I hope the MCP community pushes for user experience even if it means additional maintainer work, but it may not and that's fair too since communities have the right to set their own priorities.

[tarunanand-dev]

Hi @anuraaga - thanks for the note. I am not sure the two are mutually exclusive. A lot of Python libraries are written with Rust core. If you ask a Python developer whether their user experience is limited because of this language bindings their answer will be the opposite. For e.g. Pydantic core is written in rust, and rust is making its way into the linux kernel. As long as you can marshal the parameters and errors back and forth across the language boundary in an efficient and user friendly way it should not matter. Most users care about the SDK being ergonomic and idiomatic which is certainly possible. So net net, no user experience should be compromised with this approach however the next point to me was important.

The article on cgo is almost a decade old and things have moved on since then. The article did mention one thing that struck me as a showstopper and that was portability. I see some suggestions around wazero and that seems to be early stage. However, I would strongly recommend that for other languages like Python, Typescript, Php etc the shared core approach is better, more economical.

Ultimately you want the community to spend their shared resources most effectively and writing the same core over and over again in multiple languages doesnt sound particularly innovative. There are other interesting problems to solve!

Maybe one day when wazero or something else solves the portability problem it will be a good time to discuss if portability is very important to the golang community today.

[findleyr]

Thanks @tarunanand-dev for the idea. I think a good rule of thumb is that cgo (or other FFI) should be avoided unless it's infeasible to implement in pure Go, especially for something like an SDK which will be included in many builds. One of the reasons we think Go is a great fit for MCP is that it's a lightweight, portable, fast-compiling glue language, and any FFI would significantly reduce that value proposition.

Additionally, as of writing I don't think the implementation of the SDK is particularly difficult. That may change in the future, but right now I think the benefits of a shared implementation would not offset the downsides.

[codefromthecrypt]

I also agree that Go should be Go unless a very good reason, and I think MCP spec is simple enough and there will be enough contributors to ward off most reasons. While folks have discussed performance (and some security hints), there are many other reasons to stay in Go.

cross-compilation, coherent stack traces, easy debugging, ability to run on any platform including one without libc... all the joys folks are used in go. let's keep MCP as joyful as other key libraries in the Go ecosystem!

[jpmcb]

Thanks for the idea but, as others have said, I see no need for external bindings to Rust or C or another systems language / compiled objects.

Go's rock solid static binary, cross-compilation support makes for a good reason to not have external dependencies on system libraries or other bindings. I'd expect that if I go build an MCP server for a certain architecture or container target, it should "just work" for that architecture without having to fuss about with other targets for the bindings.

[jakemac53]

Clients support the MCP Roots feature, including roots-changed notifications. Roots can be added and removed from a Client with AddRoots and RemoveRoots:

I think this might be consistent with other MCP SDKs, but I did want to mention that having the client by default always claim to have roots support is sometimes not ideal, because while the protocol may be "supported", that doesn't mean the client actually ever will set any roots.

This makes it difficult for a server to handle clients which don't actually ever set roots, for instance they may want to conditionally add a tool to manage roots.

I ran into this developing the MCP server for Dart/Flutter tooling, cursor claims to support roots but does not actually support them, so I had to add a flag to force enable a fallback mode instead of relying on detection of the supported features.

[MegaGrindStone]

I agree with this. But, from the current implementation, it looks like the client always declare to NOT support the Roots feature. Because, in that code, the Roots field from ClientCapabilities have a zero value, or do I miss something? Maybe in that code, we could check the length of the Client's roots to decide whether this client support Roots feature or not.

[jakemac53]

Yeah my comment is based only on the design proposed, it looks to me as well like the current implementation doesn't ever declare roots support at all, but I assume that is just a work in progress, or I am missing something as well.

For dart_mcp I chose to take a more object oriented approach, where you use mixins for each feature you support. It is a bit more tricky to fit that cleanly into a functional style though - because you want some actual nice apis for adding/removing roots. The mixins can both operate sort of as middleware for handling requests and altering the capabilities, as well as adding members (add/removeRoots).

[findleyr]

@jakemac53 thanks for the feedback. We should include a way to suppress client and server capabilities.

I think it's just a bug of our type wasn't extracting root capabilities to a named type like it does for server capabilities, and therefore root capabilities were generated as an embedded struct rather than a nillable (and therefore optional) field. So @MegaGrindStone while you are right that there's something wrong with the zero value in the implementation, it's not quite what you conclude: we were indeed always advertising root support.

@jakemac53 Just checking my understanding of what you're saying: if a hypothetical client used this SDK, and yet didn't have a UI element for users to add roots, it would be buggy for servers to assume that the user has chosen not to add roots. I agree that makes sense.

On the other hand, it seems less important to be able to suppress server capabilities, but we should allow it anyway.
mcp-go adds the capability on the first call to AddXXX, which seems like a good solution.

Is there any reason to support tools (or roots) and not advertise support of listChanged? I can't imagine a scenario where that would be desirable.

[jakemac53]

Just checking my understanding of what you're saying: if a hypothetical client used this SDK, and yet didn't have a UI element for users to add roots, it would be buggy for servers to assume that the user has chosen not to add roots.

Yes, exactly

mcp-go adds the capability on the first call to AddXXX, which seems like a good solution.

This assumes at least one root will be set prior to any servers being connected - the capabilities can only be given during initialization. So, I don't think that is really a good solution.

Is there any reason to support tools (or roots) and not advertise support of listChanged? I can't imagine a scenario where that would be desirable.

I don't think there is, for dart_mcp I chose to just always support the change notifications with no opt-out.

[findleyr]

This assumes at least one root will be set prior to any servers being connected - the capabilities can only be given during initialization. So, I don't think that is really a good solution.

Yeah, I think it's necessary but perhaps not sufficient. AddTools is variadic, so AddTools() could enable the capability without actually adding any tools, but it's too clever.

[anuraaga]

Had a random thought, I was wondering what the future of the design is in terms of ownership. I found the section on package layout

Therefore, this is the core package layout, assuming github.com/modelcontextprotocol/go-sdk as the module path.

github.com/modelcontextprotocol/go-sdk/mcp: the bulk of the user facing API
github.com/modelcontextprotocol/go-sdk/jsonschema: a jsonschema implementation, with validation
github.com/modelcontextprotocol/go-sdk/internal/jsonrpc2: a fork of x/tools/internal/jsonrpc2_v2

which seems to indicate the future is in the modelcontextprotocol org. If so, why not do it as soon as possible? If that is the intention, the only reason I can think of is to want to use gerrit for code reviews, but I would suggest it's better to get used to GitHub sooner than later.

If the intention is to be golang.org/x/mcp then there doesn't seem to be a reason to change from prototyping within the tools repo, but the question then that I think should be documented is what's the intention in terms of versioning. I know many /x/ modules never release a v1, with some that do, but I guess MCP users are expecting a v1 sooner than later.

I think it would be quite helpful to clarify what the goals are in terms of ownership and community engagement in this design, probably more so than even the technical details which will all somehow be resolved before a v1 release by the community. My advice would be to add a # Community section to the doc with these or any related details that come to mind.

[findleyr]

The intention is to contribute this to github.com/modelcontextprotocol/go-sdk. I like the idea of a Community section, thanks.

[findleyr]

There's a lot of overlap between the functionality in this (generated?) comment and the existing extensibility mechanisms of the getServer factory method, SSE transport, and middleware. I certainly wouldn't want an initial version of the SDK to have two constructors for the SSEHTTPHandler: if there are missing hooks in current constructor, we should fix them.

Please provide concrete arguments for the three bullets in the 'Why this is important' section. For example, the whole point of the existing getServer factory is to provide per-session customization.

Though I do think we should consider passing options to NewSSEHandler, for future-proofing.

[jpmcb]

🔥 Design looks impeccable, as I'd expect from the Go team!!!!!! Great work Robert! Can't wait to get my hands on it

re: Governance and ownership

One area I see as missing from this discussion is how the SDK will be governed and managed - Go code is one thing, governing the Go community around the SDK is another.

From what I've inferred:

• The SDK will land in github.com/modelcontextprotocol/go-sdk
• The Go team at Google will be the majority supporters

Open questions from this structure:

• What is the onramp for new maintainers?
• Will there be a Google group or something setup for the broader community to participate in and get notifications for? (much like the golang-annouce@googlegroups.com for new version announcements or cobra-security@googlegroups.com for responsible security disclosures)
• Who owns the direction of this SDK? The Go team?

---

I don't think this is blocking, just would be good to iron out now. And maybe some of these questions are things that are better left for the core modelcontextprotocol org