TurboMCP

A ground-up Rust SDK for the Model Context Protocol — both halves of the protocol, server and client — with a macro-driven, zero-boilerplate surface and strict spec compliance as a feature.

Status: 4.0.0-alpha.5 — a prerelease for community testing. v4 is a from-scratch rewrite of TurboMCP; the stable line is 3.x. Edition 2024, MSRV 1.88. It interoperates with the official Rust SDK in both directions, on both revisions, and both halves are scored against the official MCP conformance suite: 231 successful server assertions and 488 distinct successful client scenario/check pairs using pinned client fixture corrections, with zero failures, skips, or warnings. All three advertised revisions (2025-06-18, 2025-11-25, 2026-07-28) are dated and frozen; 2026-07-28 is generated from the released schema/2026-07-28/, not the RC. Found something broken or unergonomic? Please open an issue.

• One macro defines a server. #[server] over an impl block turns #[tool] / #[resource] / #[prompt] methods into a fully-wired MCP server. The macro generates schema derivation code; schema values are initialized once at runtime and cloned for callers, and the advertised capabilities are derived from which markers are present — they can't drift from the implementation.
• Three protocol revisions, one handler. The same server answers 2025-06-18, 2025-11-25, and 2026-07-28. Your handlers speak version-neutral types; the version-specific wire shapes are conversions, not signature changes — including dropping, per session, the fields a revision predates. Pin the set with #[server(protocols("2025-11-25", …))].
• Transports behind one builder. stdio (default), Streamable HTTP (axum), and WebSocket. MyServer.run_stdio(), or .into_server().layer(…).serve(Http::bind(addr)): one runtime wires sessions, DELETE and graceful shutdown, with or without middleware. WebSocket rides the HTTP endpoint as a route, behind the same guards.
• The client too. A typed Client runs the handshake, negotiates the version, and speaks the same neutral API — interoperating with the official Rust SDK (rmcp) in both directions.
• Production seams. OAuth 2.1 on both halves (resource-server bearer validation and the client auth-code + PKCE flow), identity-keyed rate limiting, OpenTelemetry tracing + metrics, progress/logging, subscriptions, response caching (SEP-2549), and bidirectional elicitation — each opt-in behind a feature flag.

rmcp is the official SDK, maintained in the modelcontextprotocol organization. It is the reasonable default, and this project is tested against it — cross-SDK interop tests run in both directions, a TurboMCP client against an rmcp server and the reverse, on every change.

TurboMCP's interoperability tests pin rmcp 3.2. TurboMCP serves three revisions (2025-06-18, 2025-11-25, 2026-07-28) using separate generated wire types and exhaustive conversions to a neutral handler API. It does not serve 2024-11-05 or 2025-03-26.

The distinguishing APIs are macro-derived capabilities, typed per-RPC contexts, composition, caller-specific visibility, and Tower middleware. Conformance and interoperability are compatibility evidence; they do not establish performance superiority over another SDK. See deployment and migration for the limits and security contracts.

use turbomcp::prelude::*;

#[derive(Clone)]
struct Hello;

#[server(name = "hello", version = "1.0.0")]
impl Hello {
    /// Say hello to someone.
    #[tool(description = "Say hello to someone")]
    async fn hello(&self, name: String) -> McpResult<String> {
        Ok(format!("Hello, {name}!"))
    }
}

#[tokio::main]
async fn main() -> Result<(), turbomcp::ProtocolError> {
    // Logs MUST go to stderr — stdout carries the MCP protocol framing.
    Hello.run_stdio().await
}

See the turbomcp crate README for the full API tour (tools/resources/prompts, structured output, HTTP, feature flags) and the examples/.

The SDK is a Cargo workspace; the turbomcp facade re-exports the pieces most users need, so a typical dependency is just turbomcp.

Compliance is tested, not asserted:

• Official conformance suite, both halves — the @modelcontextprotocol/conformance harness runs in both directions on both scored revisions. As the server, it drives a full-featured TurboMCP server over Streamable HTTP: 236 checks, 231 pass, 0 fail, 5 informational. As the client, it stands up a deliberately awkward mock server per scenario and referees what our client did on the wire: 488 distinct successful scenario/check pairs, 0 failures, including the OAuth scenarios through the public OAuthSession coordinator. The client gate uses hash-pinned fixture corrections and reports zero failures, skips, or warnings. The unmodified upstream mode remains available with its original fixture limitations. Neither side has failure waivers (crates/turbomcp-conformance).
• Cross-SDK interop — a TurboMCP client drives an official-Rust-SDK (rmcp 3.2) server and vice-versa, in-process, on 2025-11-25 and the stateless 2026-07-28 (crates/turbomcp-interop).
• Workspace regression tests (also run against the no_std foundation configs) — dual-version dispatch, transport hardening (Origin/auth/size caps/idle reaping), handler-panic containment, MRTR elicitation, tasks (including in-execution input), subscriptions, pagination, response caching, auth negative paths, client failure semantics against misbehaving servers, and byte-level codec interchangeability (serde_json ↔ sonic-rs).
• Fuzzing + supply chain — cargo-fuzz targets for every untrusted-input decoder (JSON-RPC codec, Mcp-Param header sentinel, URI templates, and a sonic-vs-serde differential), run out of band via just fuzz; cargo-deny (advisories/bans/licenses/sources) runs in CI on every push.
• wasm-portable foundation — turbomcp-core/-codec/-protocol build no_std for wasm32-unknown-unknown on every gate run.

The macro surface is intentionally source-compatible for the common case; see crates/turbomcp/MIGRATION.md for the v3 → v4 deltas.

MIT