mcp-cpp

A modern C++20 SDK for the Model Context Protocol (MCP), targeting protocol revision 2025-11-25.

Status: beta. The wire format is locked to the official MCP spec; the C++ API may still see refinements before 1.0.

Why

LLM applications increasingly need a uniform way to talk to tool, resource, and prompt providers. MCP is that protocol; this SDK aims to be a clean, fast, spec-faithful C++ implementation suitable for embedding in editors, IDEs, agents, native applications, and high-throughput servers.

Design goals:

• Spec-faithful. Field names, method names, and behaviors match the official TypeScript schema verbatim.
• Idiomatic C++20. RAII, value semantics, std::variant for sum types, futures for async results, no inheritance hierarchies where composition will do.
• Small surface, big leverage. A focused public API; most of the work happens behind a few well-named types.
• Production-ready. Sanitizer-clean, thread-safe where it must be, no hidden allocations on hot paths.

What's in 0.1

Quick start

A minimal server

#include <mcp/mcp.hpp>

int main() {
    mcp::Server server{
        mcp::Implementation{.name = "calc", .version = "1.0.0"},
    };

    server.tool("add",
        nlohmann::json{
            {"type", "object"},
            {"properties", {
                {"a", {{"type", "number"}}},
                {"b", {{"type", "number"}}},
            }},
            {"required", nlohmann::json::array({"a", "b"})},
        },
        [](const nlohmann::json& args) -> mcp::CallToolResult {
            const double a = args.at("a").get<double>();
            const double b = args.at("b").get<double>();
            return {
                .content = { mcp::TextContent{.text = std::to_string(a + b)} },
            };
        });

    server.run(std::make_unique<mcp::StdioTransport>());
}

Run from a host that supports MCP servers (Claude Desktop, Cursor, the official inspector, etc.) by pointing at the resulting binary. Stderr is yours for diagnostics; stdout is reserved for the JSON-RPC stream.

A minimal client

#include <mcp/mcp.hpp>

int main() {
    auto pair = /* spawn the server, wire its stdio to a transport */;

    mcp::Client client{
        mcp::Implementation{.name = "my-app", .version = "0.1.0"},
    };
    client.connect(std::move(pair.transport));

    auto info = client.initialize().get();
    std::cout << "connected to " << info.server_info.name << "\n";

    auto tools = client.list_tools().get();
    for (const auto& t : tools.tools) {
        std::cout << " - " << t.name << "\n";
    }

    auto out = client.call_tool("add",
                                nlohmann::json{{"a", 2}, {"b", 3}}).get();
    if (auto* text = std::get_if<mcp::TextContent>(&out.content[0])) {
        std::cout << "result: " << text->text << "\n";
    }
}

Hosting over HTTP

Same server, exposed as a Streamable HTTP endpoint. Multiple clients each get their own Mcp-Session-Id-keyed session.

mcp::HttpServerHost host{
    mcp::Implementation{.name = "calc", .version = "1.0"},
    mcp::HttpServerHost::Options{
        .host            = "127.0.0.1",
        .port            = 8080,
        .path            = "/mcp",
        .allowed_origins = {"https://example.com"},  // DNS-rebind defense
    },
    [](mcp::Server& s) {
        s.tool("add", schema, handler);
    },
};
host.start();
// ... wait for shutdown signal ...
host.stop();

A client connects to it the same way it connects to a stdio server, but with HttpClientTransport instead of StdioTransport:

mcp::HttpClientTransport::Options topts;
topts.url = "http://localhost:8080/mcp";
mcp::Client client{mcp::Implementation{.name = "my-app", .version = "0.1"}};
client.connect(std::make_unique<mcp::HttpClientTransport>(topts));
auto info = client.initialize().get();
auto out  = client.call_tool("add", {{"a", 1}, {"b", 2}}).get();

Both examples/http_calculator_server (binds 127.0.0.1:8080 by default) and the stdio examples/calculator_server ship with the build.

A non-trivial demo: persistent notes for the assistant

examples/mcp_notes_server is a SQLite + FTS5 scratchpad that gives Claude (or any MCP client) persistent, full-text-searchable memory across sessions. Plug it into Claude Desktop / Claude Code and ask "save my Postgres prod connection string under db.prod" — it's there next time you open a chat. Search "postgres" and you get a BM25-ranked snippet hit. Glob deletes go through the spec's elicitation primitive so the human gets a yes/no prompt before anything is wiped.

Build it (libsqlite3-dev on Linux, ships with macOS):

cmake -S . -B build -DMCP_BUILD_EXAMPLES=ON
cmake --build build --target mcp_notes_server

See examples/mcp_notes_server/README.md for the wire-up.

Server-initiated LLM calls (sampling)

A common agentic pattern is for a tool to ask the host application to make an LLM call on its behalf.

server.tool("ask", schema,
    [&server](const nlohmann::json& args) -> mcp::CallToolResult {
        auto resp = server.sample(mcp::CreateMessageRequestParams{
            .messages = {
                mcp::SamplingMessage{
                    .role    = mcp::Role::user,
                    .content = mcp::TextContent{.text = args["q"]},
                },
            },
            .max_tokens = 512,
        }).get();
        return {
            .content = { mcp::TextContent{
                .text = std::get<mcp::TextContent>(resp.content).text,
            }},
        };
    });

The client plugs in a sampling handler that does the actual LLM call:

client.set_sampling_handler(
    [](const mcp::CreateMessageRequestParams& req) -> mcp::CreateMessageResult {
        // ...invoke your LLM provider with `req.messages` etc...
        return {
            .role    = mcp::Role::assistant,
            .content = mcp::TextContent{.text = "..."},
            .model   = "claude-3-5-sonnet-20241022",
        };
    });

The Session dispatches inbound requests on a worker thread, so calling server.sample(...).get() from inside a tool handler is safe — it does not deadlock.

Building

cmake -B build -G Ninja
cmake --build build
ctest --test-dir build --output-on-failure

Build options

* defaults to ON when this is the top-level project; OFF when consumed via add_subdirectory.

Consuming the library

CMake find_package

After cmake --install build, downstream projects can do:

find_package(mcp REQUIRED)
target_link_libraries(my_app PRIVATE mcp::mcp)

add_subdirectory / FetchContent

include(FetchContent)
FetchContent_Declare(mcp
    GIT_REPOSITORY https://github.com/Neumann-Labs/mcp-cpp.git
    GIT_TAG        main)
FetchContent_MakeAvailable(mcp)

target_link_libraries(my_app PRIVATE mcp::mcp)

Architecture in 60 seconds

                  +-----------------+
   application -> | Server / Client | <- public façade
                  +-----------------+
                          |
                  +-----------------+
                  |     Session     | <- JSON-RPC dispatch +
                  +-----------------+    request/response correlation
                          |
                  +-----------------+
                  |   Transport     | <- byte shovel (StdioTransport,
                  +-----------------+    Streamable HTTP, in-memory pair)
                          |
                          v
                       wire

• Transport is a small abstract interface: it shovels frames; the Session parses them.
• Session owns the request-id generator, the pending-request map, the request/notification dispatch table, and the timeout sweep. It's the same class on both sides of the wire — Server and Client just register different handlers and call different convenience methods.
• Server / Client are thin façades: they translate typed C++ values into JSON-RPC frames and back, expose tool() / resource() / prompt() / sample() / list_tools() / etc., and own one Session each.

Threading

• The transport reads on its own thread and invokes the Session's on-message callback.
• Inbound requests are dispatched to a detached worker thread so user handlers can themselves issue further requests on the same Session (e.g. server.sample(...).get() from inside a tool handler). Session::close() waits for those workers to finish before tearing members down.
• Inbound notifications and responses run on the read thread directly.
• send_request / send_notification are safe to call from any thread.

License

Apache 2.0 — see LICENSE.