VM
5 stars 4 forks Quality 60 Trend 26Apache-2.0

MCP is a Go implementation of the Model Context Protocol — a standardized way for applications to communicate with AI models.

Overview

MCP is a Go implementation of the Model Context Protocol — a standardized way for applications to communicate with AI models.

README

MCP (Model Context Protocol) for Go

MCP is a Go implementation of the Model Context Protocol — a standardized way for applications to communicate with AI models. It allows developers to seamlessly bridge applications and AI models using a lightweight, JSON-RPC–based protocol.

This repository provides the Go MCP client/server runtime, bridge, and auth helpers. Protocol definitions and schemas live in mcp-protocol.

Official Model Context Protocol Specification

Overview

MCP (Model Context Protocol) is designed to provide a standardized communication layer between applications and AI models. The protocol simplifies the integration of AI capabilities into applications by offering a consistent interface for resource access, prompt management, model interaction, and tool invocation.

Key features:

  • JSON-RPC 2.0–based communication
  • Support for multiple transport protocols (HTTP/SSE, stdio)
  • Server-side features:
    • Resource management
    • Model prompting and completion
    • Tool invocation
    • Subscriptions for resource updates
    • Logging
    • Progress reporting
    • Request cancellation
  • Client-side features:
    • Roots
    • Sampling
    • Elicitation

Architecture

For detailed guides on custom server implementations and authentication, see docs/implementer.md and docs/authentication.md.

MCP is built around the following components:

  1. Transport: JSON-RPC over HTTP/SSE, streamable HTTP, or stdio
  2. Server: Handles incoming requests and dispatches to the protocol implementation
  3. Client: Makes requests to MCP-compatible servers
  4. Protocol Implementation (server.Handler): Provides the actual functionality behind each protocol method

High-Level Architecture

graph LR
    Client[MCP Client] -->|JSON-RPC / HTTP/SSE| Server[MCP Server]
    Server -->|Dispatches to| Handler[MCP Server]
    subgraph Auth[Authentication / Authorization]
      OAuth2[OAuth2 / OIDC]
    end
    Client -.->|Bearer Token| OAuth2
    Server -.->|Token Validation| OAuth2

Getting Started

Installation

go get github.com/viant/mcp

If you just need to connect existing tools to a remote MCP server you might prefer to use the standalone Bridge binary instead of embedding the Go package. See Bridge Guide for details.

Creating a Server

Quick Start: Default Server

DefaultHandler can be used to quickly set up an MCP server. It provides no-op stubs for all methods, allowing you to focus on implementing only the methods you need. Register handlers inline without writing a custom server type:

package main

import (
  "context"
  "encoding/json"
  "fmt"
  "github.com/viant/jsonrpc"
  "github.com/viant/mcp-protocol/schema"
  serverproto "github.com/viant/mcp-protocol/server"
  "github.com/viant/mcp/server"
  "log"
)

func main() {

  type Addition struct {
    A int `json:"a"`
    B int `json:"b"`
  }

  type Output struct {
    Result int
  }

  newHandler := serverproto.WithDefaultHandler(context.Background(), func(server *serverproto.DefaultHandler) error {
    // Register a simple resource
    server.RegisterResource(schema.Resource{Name: "hello", Uri: "/hello"},
      func(ctx context.Context, request *schema.ReadResourceRequest) (*schema.ReadResourceResult, *jsonrpc.Error) {
        return &schema.ReadResourceResult{Contents: []schema.ReadResourceResultContentsElem{{Text: "Hello, world!"}}}, nil
      })

    type Addition struct {
      A int `json:"a"`
      B int `json:"b"`
    }

    type Result struct {
      Result int `json:"acc"`
    }
    // Register a simple calculator tool: adds two integers
    if err := serverproto.RegisterTool[*Addition, *Result](server.Registry, "add", "Add two integers", func(ctx context.Context, input *Addition) (*schema.CallToolResult, *jsonrpc.Error) {
      sum := input.A + input.B
      out := &Result{Result: sum}
      data, err := json.Marshal(out)
      if err != nil {
        return nil, jsonrpc.NewInternalError(fmt.Sprintf("failed to marshal result: %v", err), nil)
      }
      return &schema.CallToolResult{Content: []schema.CallToolResultContentElem{{Text: string(data)}}}, nil
    }); err != nil {
      return err
    }
    return nil
  })

  srv, err := server.New(
    server.WithNewHandler(newHandler),
    server.WithImplementation(schema.Implementation{Name: "default", Version: "1.0"}),
  )
  if err != nil {
    log.Fatalf("Failed to create server: %v", err)
  }

  // Choose one transport (see below). Example: HTTP (SSE by default)
  log.Fatal(srv.HTTP(context.Background(), ":4981").ListenAndServe())

}

Starting the Server

  • Stdio (typical for editor integrations):

    stdioSrv := srv.Stdio(context.Background())
    log.Fatal(stdioSrv.ListenAndServe())
    
  • HTTP over SSE (default):

    httpSrv := srv.HTTP(context.Background(), ":4981")
    log.Fatal(httpSrv.ListenAndServe())
    
  • HTTP Streamable (toggle):

    srv.UseStreamableHTTP(true)
    httpSrv := srv.HTTP(context.Background(), ":4981")
    log.Fatal(httpSrv.ListenAndServe())
    

If you prefer a single configuration object, see NewServer in server.go which accepts transport options and can enable Streamable HTTP based on Transport.Type.

Endpoints and Routing

  • By default, both transports are mounted:
    • SSE: GET /sse and POST /message
    • Streamable HTTP: POST/GET /mcp
  • You can enable root redirect so "/" forwards to the active transport base:
    • WithRootRedirect(true)
  • You can also customize the endpoint paths:
    • WithStreamableURI("/api/mcp")
    • WithSSEURI("/api/sse")
    • WithSSEMessageURI("/api/rpc")

Example:

srv, _ := mcp.New(
  mcp.WithNewHandler(newHandler),
  mcp.WithStreamableURI("/api/mcp"),
  mcp.WithSSEURI("/api/sse"),
  mcp.WithSSEMessageURI("/api/rpc"),
  mcp.WithRootRedirect(true), // redirects "/" to /api/mcp if UseStreamableHTTP(true), else /api/sse
)
srv.UseStreamableHTTP(true)
httpSrv := srv.HTTP(context.Background(), ":4981")
log.Fatal(httpSrv.ListenAndServe())

Add a Resource

Register a readable resource URI and return its content from your handler.

h.RegisterResource(schema.Resource{Name: "hello", Uri: "/hello"},
  func(ctx context.Context, req *schema.ReadResourceRequest) (*schema.ReadResourceResult, *jsonrpc.Error) {
    return &schema.ReadResourceResult{Contents: []schema.ReadResourceResultContentsElem{{Text: "Hello, world!", Uri: req.Params.Uri}}}, nil
  })

To notify clients about updates, emit resources/updated via h.Notifier.

Git-backed Resources

Expose curated repository files by plugging in the resource/git helper:

import resgit "github.com/viant/mcp/resource/git"

cfg := &resgit.Config{
  Repos: []resgit.Repo{
    {
      ID:                "mcp",
      URL:               "https://github.com/viant/mcp.git",
      Roots:             []string{"docs"},
      Include:           []string{"**/*.md"},
      ReadMeOnly:        true,
      ExposeSnapshotZip: true,
    },
  },
}

proto.WithDefaultHandler(ctx, func(h *serverproto.DefaultHandler) error {
  if err := resgit.Register(ctx, h, cfg); err != nil {
    return err
  }
  return nil
})

Clients can now call resources/list to discover URIs like repo://mcp/docs/server_guide.md and resources/read or repo://mcp/snapshot.zip to fetch either trimmed files or the exact snapshot zip (MD5 included via _meta).

Add a Prompt

Expose reusable prompt templates that clients can list and resolve.

prompt := &schema.Prompt{Name: "welcome", Arguments: []schema.PromptArgument{{Name: "name", Required: ptr(true)}}}
h.RegisterPrompts(prompt, func(ctx context.Context, p *schema.GetPromptRequestParams) (*schema.GetPromptResult, *jsonrpc.Error) {
  return &schema.GetPromptResult{Messages: []schema.PromptMessage{{Role: schema.RoleAssistant, Content: schema.TextContent{Type: "text", Text: "Hello, " + p.Arguments["name"] + "!"}}}}, nil
})

See the Server Guide for deeper coverage of resources and prompts.

Further Reading

Creating a Client

You can connect to an MCP server over stdio, HTTP/SSE, or HTTP streamable.

  • SSE client (simple):

    ctx := context.Background()
    sseTransport, _ := sse.New(ctx, "http://localhost:4981/sse")
    cli := client.New("Demo", "1.0", sseTransport)
    if _, err := cli.Initialize(ctx); err != nil { log.Fatal(err) }
    tools, _ := cli.ListTools(ctx, nil)
    fmt.Println("tools:", len(tools.Tools))
    
  • Streamable client:

    ctx := context.Background()
    streamTransport, _ := streamable.New(ctx, "http://localhost:4981/")
    cli := client.New("Demo", "1.0", streamTransport)
    _, _ = cli.Initialize(ctx)
    
  • Stdio client (spawn a child process):

    stdioTransport, _ := stdio.New("./your-mcp-server-binary",
      stdio.WithArguments("--flag1", "value"))
    cli := client.New("Demo", "1.0", stdioTransport)
    _, _ = cli.Initialize(context.Background())
    
  • OAuth2/OIDC (optional):

    See the Authentication section below for creating an http.Client with token handling and passing it to the SSE/streamable transports.

Advanced: If you need automatic reconnect and integrated auth, use the helper mcp.NewClient(handler, *ClientOptions) from the root package, which builds the transport and wires an auth interceptor. Supply a pclient.Handler implementation if your client needs to support server-initiated calls (roots, sampling, elicitation).

Authentication & Authorization Summary

MCP supports transport-agnostic authentication and authorization (HTTP or HTTP-SSE) via OAuth2/OIDC in two modes:

  • Global Resource Protection (spec-based):

    • github.com/viant/mcp/server/auth.AuthServer enforces a Bearer token across all endpoints, except those excluded via ExcludeURI (e.g. /sse).
    • Configure by creating an auth.Service from authorization.Policy and wiring it with: server.WithProtectedResourcesHandler(service.ProtectedResourcesHandler), server.WithAuthorizer(service.Middleware), and server.WithJRPCAuthorizer(service.EnsureAuthorized).
    • Exposes /.well-known/oauth-protected-resource for metadata discovery (RFC 9728).
  • Fine-Grained Tool/Resource Control (experimental):

    • Implements auth.Authorizer in AuthServer.EnsureAuthorized, returning 401 Unauthorized per JSON-RPC request.
    • Configure per-tool/resource metadata via Config.Tools or Config.Tenants in authorization.Policy.
  • Fallback Token Fetching:

    • github.com/viant/mcp/server/auth.FallbackAuth wraps a strict AuthServer.
    • On 401 challenge, fetches tokens via ProtectedResourceTokenSource and optional ID tokens via IdTokenSource, then retries.
    • Create with auth.NewFallbackAuth(strictAuthServer, tokenSource, idTokenSource).
  • Client-Side Support:

    • Use github.com/viant/mcp/client/auth/transport.New with:
      • WithStore(store.Store): handles client config, metadata & token caching.
      • WithAuthFlow(flow.AuthFlow): selects interactive auth flow (e.g., browser PKCE).
    • The RoundTripper automatically:
      1. Handles the initial 401 challenge (WWW-Authenticate header).
      2. Discovers protected resource and auth server metadata.
      3. Acquires tokens and retries the original request with Authorization: Bearer .
  • Example SSE integration:

    rt, _ := transport.New(
        transport.WithStore(myStore),
        transport.WithAuthFlow(flow.NewBrowserFlow()),
    )
    httpClient := &http.Client{Transport: rt}
    sseTransport, _ := sse.New(ctx, "https://myapp.example.com/sse", sse.WithHttpClient(httpClient))
    mcpClient := client.New("MyClient", "1.0", sseTransport, client.WithCapabilities(schema.ClientCapabilities{}))
    

These features are transport-agnostic and apply equally over HTTP, SSE, or other supported transports.

Protocol Methods

MCP supports the following Server Side methods:

  • initialize - Initialize the connection
  • ping - Check server status
  • resources/list - List available resources
  • resources/read - Read resource contents
  • resources/templates/list - List resource templates
  • resources/subscribe - Subscribe to resource updates
  • resources/unsubscribe - Unsubscribe from resource updates
  • prompts/list - List available prompts
  • prompts/get - Get prompt details
  • tools/list - List available tools
  • tools/call - Call a specific tool
  • complete - Get completions from the model
  • logging/setLevel - Set logging level

MCP supports the following Client Side methods:

  • roots/list - List available roots
    • sampling/createMessage - A standardized way for servers to request LLM sampling (“completions” or “generations”) from language models via clients.
    • elicitation/create - Ask the client to collect additional user input (forms or external links) and return structured content.
    • interaction/create - Initiate a user-interaction request, allowing the server to prompt the user via the client UI.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the Apache License 2.0.

Credits

Author: Adrian Witas

This project is maintained by Viant.

Tool-call JSON numbers

Generic tool-call argument and result values use standard encoding/json decoding: numeric leaves are float64, including large integers and precise decimals. Large numbers may round, tiny exponents may underflow to zero, and float64 overflow is rejected. Trailing or malformed JSON is rejected.

Declared numeric fields and envelope IDs retain their existing typed decoding. Schema enums and schema vocabulary continue to preserve exact json.Number values; this exact schema handling does not extend to generic tool payloads. Applications requiring exact business identifiers should use string fields.

View the full README on GitHub

Install

This server does not publish a one-line install command.

Open the repository installation guide New to MCP? Start with the mcpskill.org guide →