Implements the Model Context Protocol, enabling client and server integration with LLM surfaces. Supports standard transports, handles protocol messages, and offers sample setups for server and client configurations.
Kotlin Multiplatform SDK for the Model Context Protocol. It enables Kotlin applications targeting JVM, Native, JS, and Wasm to implement MCP clients and servers using a standardized protocol interface.
The Model Context Protocol allows applications to provide context for LLMs in a standardized way, separating the concerns of providing context from the actual LLM interaction. This Kotlin SDK implements the MCP specification, making it easy to:
io.modelcontextprotocol:kotlin-sdk – umbrella SDK (client + server APIs)io.modelcontextprotocol:kotlin-sdk-client – client-only APIsio.modelcontextprotocol:kotlin-sdk-server – server-only APIsAdd the Maven Central repository and the SDK dependency:
repositories {
mavenCentral()
}
dependencies {
// See the badge above for the latest version
implementation("io.modelcontextprotocol:kotlin-sdk:$mcpVersion")
}
Use kotlin-sdk-client or kotlin-sdk-server if you only need one side of the API:
dependencies {
implementation("io.modelcontextprotocol:kotlin-sdk-client:$mcpVersion")
implementation("io.modelcontextprotocol:kotlin-sdk-server:$mcpVersion")
}
In a Kotlin Multiplatform project you can add the SDK to commonMain:
commonMain {
dependencies {
// Works as a common dependency as well as the platform one
implementation("io.modelcontextprotocol:kotlin-sdk:$mcpVersion")
}
}
The Kotlin MCP SDK uses Ktor, but it does not add Ktor engine dependencies transitively. You need to declare Ktor client/server dependencies yourself (or reuse the ones already used in your project), for example:
dependencies {
// MCP client with Ktor
implementation("io.ktor:ktor-client-cio:$ktorVersion")
implementation("io.modelcontextprotocol:kotlin-sdk-client:$mcpVersion")
// MCP server with Ktor
implementation("io.ktor:ktor-server-netty:$ktorVersion")
implementation("io.modelcontextprotocol:kotlin-sdk-server:$mcpVersion")
}
Let's create a simple MCP client and server to demonstrate the basic usage of the Kotlin SDK.
Create an MCP client that connects to a server via Streamable HTTP transport and lists available tools:
io.ktor.client.HttpClient
io.ktor.client.plugins.sse.SSE
io.modelcontextprotocol.kotlin.sdk.client.Client
io.modelcontextprotocol.kotlin.sdk.client.StreamableHttpClientTransport
io.modelcontextprotocol.kotlin.sdk.types.Implementation
kotlinx.coroutines.runBlocking
= runBlocking {
url = args.firstOrNull() ?:
httpClient = HttpClient { install(SSE) }
client = Client(
clientInfo = Implementation(
name = ,
version =
)
)
transport = StreamableHttpClientTransport(
client = httpClient,
url = url
)
client.connect(transport)
tools = client.listTools().tools
println(tools)
}
Create an MCP server that exposes a simple tool and runs on an embedded Ktor server with Streamable HTTP transport. For a full working project with all required dependencies, see the simple-streamable-server sample.
You can run the server and then connect to it using the client or test with the MCP Inspector:
npx -y @modelcontextprotocol/inspector
In the inspector UI, connect to http://localhost:3000/mcp.
The MCP protocol defines core primitives that enable communication between servers and clients:
Capabilities define what features a server or client supports. They are declared during initialization and determine what operations are available.
Servers declare their capabilities to inform clients what features they provide:
Clients declare their capabilities to inform servers what features they support:
The Server API lets you wire prompts, resources, and tools with only a few lines of Kotlin. Each feature is registered
up front and then resolved lazily when a client asks for it, so your handlers stay small and suspendable.
Prompts are user-controlled templates that clients discover via prompts/list and fetch with prompts/get when a user
chooses one (think slash commands or saved flows). They’re best for repeatable, structured starters rather than ad-hoc
model calls.
val server = Server(
serverInfo = Implementation(
name = "example-server",
version = "1.0.0"
),
options = ServerOptions(
capabilities = ServerCapabilities(
prompts = ServerCapabilities.Prompts(listChanged = true),
),
)
)
server.addPrompt(
name = "code-review",
description = "Ask the model to review a diff",
arguments = listOf(
PromptArgument(name = "diff", description = "Unified diff", required = ),
),
) { request ->
GetPromptResult(
description = ,
messages = listOf(
PromptMessage(
role = Role.User,
content = TextContent(text = ),
),
),
)
}
Use prompts for anything that deserves a template: bug triage questions, onboarding checklists, or saved searches. Set
listChanged = true only if your prompt catalog can change at runtime and your server will emit
notifications/prompts/list_changed when it does.
Resources are application-driven context that clients discover via resources/list or resources/templates/list, then
fetch with resources/read. Register each one with a stable URI and return a ReadResourceResult when asked—contents
can be text or binary blobs.
val server = Server(
serverInfo = Implementation(
name = "example-server",
version = "1.0.0"
),
options = ServerOptions(
capabilities = ServerCapabilities(
resources = ServerCapabilities.Resources(subscribe = true, listChanged = true),
),
)
)
server.addResource(
uri = "note://release/latest",
name = "Release notes",
description = "Last deployment summary",
mimeType = "text/markdown",
) { request ->
ReadResourceResult(
contents = listOf(
TextResourceContents(
text = "Ship 42 reached production successfully.",
uri = request.uri,
mimeType = ,
),
),
)
}
Resources can be static text, generated JSON, or blobs—anything the client can surface to the user or inject into the
model context. Set subscribe = true if you emit notifications/resources/updated for changes to specific URIs, and
listChanged = true if you’ll send notifications/resources/list_changed when the catalog itself changes.
Tools are model-controlled capabilities the client exposes to the model. Clients discover them via tools/list, invoke
them with tools/call, and your handlers receive JSON arguments, can emit streaming logs or progress, and return a
CallToolResult. Keep a human in the loop for sensitive operations.
val server = Server(
serverInfo = Implementation(
name = "example-server",
version = "1.0.0"
),
options = ServerOptions(
capabilities = ServerCapabilities(
tools = ServerCapabilities.Tools(listChanged = true),
),
)
)
server.addTool(
name = "echo",
description = "Return whatever the user sent back to them",
) { request ->
val text = request.arguments?.get("text")?.jsonPrimitive?.content ?: "(empty)"
CallToolResult(content = listOf(TextContent(text = "Echo: $text")))
}
Register as many tools as you need—long-running jobs can report progress via the request context, and tools can also
trigger sampling (see below) when they need the client’s LLM. Set listChanged = true only if your tool catalog can
change at runtime and your server will emit notifications/tools/list_changed when it does.
Completion provides argument suggestions for prompts or resource templates. Declare the completions capability and
handle completion/complete requests to return up to 100 ranked values (include total/hasMore if you paginate).
val server = Server(
serverInfo = Implementation(
name = "example-server",
version = "1.0.0"
),
options = ServerOptions(
capabilities = ServerCapabilities(
completions = ServerCapabilities.Completions,
),
)
)
val session = server.createSession(
StdioServerTransport(
inputStream = System.`in`.asSource().buffered(),
outputStream = System.out.asSink().buffered()
)
)
session.setRequestHandler<CompleteRequest>(Method.Defined.CompletionComplete) { request, _ ->
options = listOf(, , )
matches = options.filter { it.startsWith(request.argument.value.lowercase()) }
CompleteResult(
completion = CompleteResult.Completion(
values = matches.take(),
total = matches.size,
hasMore = matches.size > ,
),
)
}
Use context.arguments to refine suggestions for dependent fields (e.g., framework list filtered by chosen language).
Logging lets the server stream structured log notifications to the client using RFC 5424 levels (debug → emergency).
Declare the logging capability; clients can raise the minimum level with logging/setLevel, and the server emits
notifications/message with severity, optional logger name, and JSON data.
val server = Server(
serverInfo = Implementation("example-server", "1.0.0"),
options = ServerOptions(
capabilities = ServerCapabilities(
logging = ServerCapabilities.Logging,
),
)
)
val session = server.createSession(
StdioServerTransport(
inputStream = System.`in`.asSource().buffered(),
outputStream = System.out.asSink().buffered()
)
)
session.sendLoggingMessage(
LoggingMessageNotification(
LoggingMessageNotificationParams(
level = LoggingLevel.Info,
logger = "startup",
data = buildJsonObject { put("message", "Server started") },
),
),
)
Keep logs free of sensitive data, and expect clients to surface them in their own UI.
List operations return paginated results with an opaque nextCursor, clients echo that cursor to fetch the next page.
Supported list calls: resources/list, resources/templates/list, prompts/list, and tools/list.
Treat cursors as opaque—don’t parse or persist them across sessions.
Include nextCursor only when more items remain an absent cursor ends pagination.
Clients advertise their capabilities (roots, sampling, elicitation, etc.) during initialization. After that they can
serve requests from the server while still initiating calls such as listTools or callTool.
Roots let the client declare where the server is allowed to operate. Declare the roots capability, respond to
roots/list, and emit notifications/roots/list_changed if you set listChanged = true. URIs must be file://
paths.
val client = Client(
clientInfo = Implementation("demo-client", "1.0.0"),
options = ClientOptions(
capabilities = ClientCapabilities(roots = ClientCapabilities.Roots(listChanged = true)),
),
)
client.addRoot(
uri = "file:///Users/demo/projects",
name = "Projects",
)
client.sendRootsListChanged()
Call addRoot/removeRoot whenever your file system view changes, and use sendRootsListChanged() to notify the
server. Keep root lists user-controlled and revoke entries that are no longer authorized.
Sampling lets the server ask the client to call its preferred LLM. Declare the sampling capability (and
sampling.tools if you allow tool-enabled sampling), and handle sampling/createMessage. Keep a human in the loop for
approvals.
val client = Client(
clientInfo = Implementation("demo-client", "1.0.0"),
options = ClientOptions(
capabilities = ClientCapabilities(
sampling = buildJsonObject { putJsonObject("tools") { } }, // drop tools if you don't support tool use
),
),
)
client.setRequestHandler<CreateMessageRequest>(Method.Defined.SamplingCreateMessage) { request, _ ->
val content = request.messages.lastOrNull()?.content
val prompt = if (content TextContent) content.text
CreateMessageResult(
model = ,
role = Role.Assistant,
content = TextContent(text = ),
)
}
Inside the handler you can pick any model/provider, require approvals, or reject the request. If you don’t support tool
use, omit sampling.tools from capabilities.
All transports share the same API surface, so you can change deployment style without touching business logic. Pick the transport that best matches where the server runs.
StdioClientTransport and StdioServerTransport tunnel MCP messages over stdin/stdout—perfect for editor plugins or
CLI tooling that spawns a helper process. No networking setup is required.
StreamableHttpClientTransport and the Ktor mcpStreamableHttp() / mcpStatelessStreamableHttp() helpers expose MCP
over a single HTTP endpoint with optional JSON-only or SSE streaming responses. This is the recommended choice for
remote deployments and integrates nicely with proxies or service meshes.
These helpers automatically install ContentNegotiation with McpJson — do not install it yourself, or a warning will
be logged. Both accept a path parameter (default: "/mcp") to mount the endpoint at any URL:
embeddedServer(CIO, port = 3000) {
mcpStreamableHttp(path = "/api/mcp") {
MyServer()
}
}.start(wait = true)
CORS for browser-based clients (e.g. MCP Inspector): if you connect from a browser-based client you need to install the Ktor CORS plugin so that MCP-specific headers are allowed and exposed:
install(CORS) {
anyHost() // restrict to specific origins in production
allowMethod(HttpMethod.Options)
allowMethod(HttpMethod.Get)
allowMethod(HttpMethod.Post)
allowMethod(HttpMethod.Delete)
allowNonSimpleContentTypes = true
allowHeader("Mcp-Session-Id")
allowHeader("Mcp-Protocol-Version")
exposeHeader("Mcp-Session-Id")
exposeHeader("Mcp-Protocol-Version")
}
Server-Sent Events remain available for backwards compatibility with older MCP clients. Two Ktor helpers are provided:
Application.mcp { } — installs SSE and ContentNegotiation with McpJson automatically, then registers MCP
endpoints at /. Do not install ContentNegotiation yourself — the SDK handles it.embeddedServer(CIO, port = 3000) {
install(SSE)
routing {
route("/api/mcp") {
mcp { MyServer() }
}
}
}.start(wait = true)
Prefer Streamable HTTP for new projects.
WebSocketClientTransport plus the matching server utilities provide full-duplex, low-latency connections—useful when
you expect lots of notifications or long-running sessions behind a reverse proxy that already terminates WebSockets.
ChannelTransport provides a simple, non-networked transport for testing and local development.
It uses Kotlin coroutines channels to provide a full-duplex connection between a client and server,
allowing for easy testing of MCP functionality without the need for network setup.
Start a sample HTTP server on port 3000:
./gradlew :samples:kotlin-mcp-server:run
Connect with the MCP Inspector or Claude Desktop/Code:
The samples directory contains runnable projects demonstrating MCP server and client implementations with various transports. See the samples overview for a comparison table and detailed descriptions.
Please see the contribution guide and the Code of conduct before contributing.
This project is licensed under Apache 2.0 for new contributions, with existing code under MIT—see the LICENSE file for details.
import io.ktor.server.cio.CIO
import io.ktor.server.engine.embeddedServer
import io.modelcontextprotocol.kotlin.sdk.server.Server
import io.modelcontextprotocol.kotlin.sdk.server.ServerOptions
import io.modelcontextprotocol.kotlin.sdk.server.mcpStreamableHttp
import io.modelcontextprotocol.kotlin.sdk.types.CallToolResult
import io.modelcontextprotocol.kotlin.sdk.types.Implementation
import io.modelcontextprotocol.kotlin.sdk.types.ServerCapabilities
import io.modelcontextprotocol.kotlin.sdk.types.TextContent
import io.modelcontextprotocol.kotlin.sdk.types.ToolSchema
import kotlinx.serialization.json.buildJsonObject
import kotlinx.serialization.json.put
fun main(args: Array<String>) {
val port = args.firstOrNull()?.toIntOrNull() ?: 3000
val mcpServer = Server(
serverInfo = Implementation(
name = "example-server",
version = "1.0.0"
),
options = ServerOptions(
capabilities = ServerCapabilities(
tools = ServerCapabilities.Tools(listChanged = true),
),
)
)
mcpServer.addTool(
name = "example-tool",
description = "An example tool",
inputSchema = ToolSchema(
properties = buildJsonObject {
put("input", buildJsonObject { put("type", "string") })
}
)
) { request ->
CallToolResult(content = listOf(TextContent("Hello, world!")))
}
embeddedServer(CIO, host = "127.0.0.1", port = port) {
mcpStreamableHttp {
mcpServer
}
}.start(wait = true)
}
| Primitive | Server Role | Client Role | Description |
|---|
| Prompts | Provides prompt templates with optional arguments | Requests and uses prompts | Interactive templates for LLM interactions |
| Resources | Exposes data sources (files, API responses, etc.) | Reads and subscribes to resources | Contextual data for augmenting LLM context |
| Tools | Defines executable functions | Calls tools to perform actions | Functions the LLM can invoke to take actions |
| Sampling | Requests LLM completions from client | Executes LLM calls and returns results | Server-initiated LLM requests (reverse direction) |
| Capability | Feature Flags | Description |
|---|
prompts | listChanged | Prompt template management and notifications |
resources | subscribelistChanged | Resource exposure, subscriptions, and update notifications |
tools | listChanged | Tool discovery, execution, and list change notifications |
logging | - | Server logging to client console |
completions | - | Argument autocompletion suggestions |
experimental | Custom properties | Non-standard experimental features |
| Capability | Feature Flags | Description |
|---|
sampling | - | Client can sample from an LLM (execute model requests) |
roots | listChanged | Client exposes root directories and can notify of changes |
elicitation | - | Client can display schema/form dialogs for structured input |
experimental | Custom properties | Non-standard experimental features |
val server = Server(
serverInfo = Implementation("example-server", "1.0.0"),
options = ServerOptions(
capabilities = ServerCapabilities(
resources = ServerCapabilities.Resources(),
),
)
)
val session = server.createSession(
StdioServerTransport(
inputStream = System.`in`.asSource().buffered(),
outputStream = System.out.asSink().buffered()
)
)
val resources = listOf(
Resource(uri = "note://1", name = "Note 1", description = "First"),
Resource(uri = "note://2", name = "Note 2", description = "Second"),
Resource(uri = "note://3", name = "Note 3", description = "Third"),
)
val pageSize = 2
session.setRequestHandler<ListResourcesRequest>(Method.Defined.ResourcesList) { request, _ ->
val start = request.params?.cursor?.toIntOrNull() ?: 0
val page = resources.drop(start).take(pageSize)
val next = if (start + page.size < resources.size) (start + page.size).toString() else null
ListResourcesResult(
resources = page,
nextCursor = next,
)
}
Route.mcp { } — registers MCP endpoints at the current route path; requires install(SSE) in the application
first. Use this to host MCP alongside other routes or under a path prefix:npx -y @modelcontextprotocol/inspector --connect http://localhost:3000
# or
claude mcp add --transport http kotlin-mcp http://localhost:3000
In the Inspector, confirm prompts, tools, resources, completions, and logs show up. Iterate locally until you’re ready to host the server wherever you prefer.
Surfaced from shared tags and platforms — no rankings paid for.