Skip to content
← Blog

Release

TurboMCP v3.6.0: The Current Release of the Supported v3 Line

TurboMCP 3.6.0 fixes Client::connect_tcp, dropped TCP and Unix messages, three security gaps, and trims a default build from 199 to 175 crates. v3 stays the maintained line while v4 is in alpha.

TurboMCP Engineering Team · · 8 min read

TurboMCP 3.6.0 (2026-09-30) is the current release of the v3 line. v3 is the maintained, supported line for production use today. TurboMCP v4 is in alpha on the main branch, and we cover it in where v4 stands today. This release is a stabilisation and hardening release for v3: it fixes a transport bug that shipped in 3.5.0, closes three security gaps, and makes the dependency tree smaller. No public API was removed and no feature was removed.

If you are on 3.5.0 and use Client::connect_tcp, upgrade now.

Why this release exists

Client::connect_tcp failed against every server in 3.5.0. The TCP client registered its connection in a task that had not yet run when initialize went out (#60). Nothing in the test suite ran a real client against a real server over TCP, so nothing caught it.

A new end-to-end test, transport_e2e, now does, over TCP, Unix, HTTP and WebSocket. A diligence pass before release then found more problems, summarised below.

Transport fixes

Problem What changed
Client::connect_tcp failed against every server (introduced in 3.5.0) TcpTransport’s client connect registered the connection from inside the task it spawned, so a send straight after connect failed with “No active TCP connections”. The connection is now registered before connect returns. The server side registers accepted connections the same way, so the connection limit counts a connection from the moment it is accepted.
TCP and Unix transports dropped messages when a peer fell behind send queued without waiting and, once a connection’s queue of 100 was full, logged a warning, dropped the message and returned Ok. The request then waited for a response that could never come. The readers did the same with incoming messages. Both now wait for room, so a slow peer slows the sender instead, and a send that reaches no connection returns an error.
HTTP graceful shutdown ran backwards On SIGINT or SIGTERM, the with_graceful_shutdown duration was spent before shutdown began (capped at 60 seconds) with the server still accepting connections, and the drain that followed had no deadline. A client holding a listening GET stream kept it open indefinitely. The server now stops accepting connections at once, ends listening GET streams, and waits up to the configured duration for in-flight requests. Without a duration it waits for them as long as they take, as before.
A request could run twice after its session expired When the server answered 404 while the client was resuming a request’s SSE stream, the client started a new session and re-sent a request the server had already received. That request now fails with ConnectionLost, and the expiry is reported to the next request before it is sent, so a new session still starts automatically. A 404 on the POST itself, which the server never ran, is still retried.
Streamable HTTP answered expired sessions with 410 The spec requires 404, which is the status a client starts a new session on. StreamableError::SessionExpired and SessionTerminated now convert to 404.
StdioTransport reported 1 MiB max message size while accepting 10 MiB It now reports the limit it enforces (limits.max_response_size, 10 MiB by default), as TCP and Unix do.

Client and auth fixes

  • tasks/cancel is capability-gated. 3.5.0 began cancelling the task an abandoned task-augmented call creates without checking that the server declared tasks.cancel, and Client::cancel_task never checked either. Without the capability an abandoned task is now left to run, and cancel_task fails without sending anything, like the other capability-gated calls.
  • FetcherConfig’s timeout and user agent apply again. In 3.5.0 the authorization-server discovery and client-metadata fetchers moved onto the SSRF validator’s DNS-pinned client, which carried the SSRF policy’s timeout and no user agent. Both are now set on each request.
  • OAuth2Client::client_credentials_flow sends resource. 3.5.0 added RFC 8707 resource indicators to the authorization-code, code exchange and refresh requests, as MCP’s authorization spec requires, and missed this grant. An authorization server that sets a token’s audience from resource could issue a token the MCP server then rejected.

Other fixes

  • WASM #[prompt] with arguments now compiles. The generated registration wrapped the arguments type in a second Option, so the documented args: Option<T> failed with a type mismatch, and so did a bare T. Option<T> works now, and a bare T produces an error telling you to take Option<T>, since a client may call a prompt without arguments.
  • TelemetryConfig::prometheus_port starts a listener. The exporter called install_recorder, which installs the recorder and nothing else, so scrapes were refused while the log said the endpoint had started. The listener now runs on the current Tokio runtime, or on its own thread outside one. The exporter answers on any path, so prometheus_path is informational, and its docs now say so.
  • Documentation. ClientBuilder::with_max_retries and with_retry_delay only affect build_resilient. build and build_sync never retry. The methods, the ConnectionConfig fields and the example now say so.

Security

Three gaps are closed.

  1. JwksClient with an SSRF validator could be steered by DNS rebinding. It validated the JWKS URI and then let its own client resolve the host again to connect, the same gap 3.5.0 closed for the discovery and client-metadata fetchers. It now connects to the address the validator checked and still never follows a redirect. Its 64 KB response limit was previously checked only after the whole body had been read into memory, and is now enforced as the body arrives.
  2. SsrfPolicy { allow_redirects: true, .. } followed redirects anywhere. The client from SsrfValidator::create_pinned_client checked only the first URL, then followed up to max_redirects hops to any host, including private and cloud-metadata addresses the policy refuses. Each hop is now validated against the same policy, with its host resolved and every address checked, before it is followed. Redirects remain off by default.
  3. ChildProcessTransport buffered a child’s stdout line in full before checking its size. A child that wrote without newlines grew the parent’s memory without bound. Lines are now read holding at most max_message_size bytes. An oversized response is still answered with an error for its id, found in the first 4 KiB of the line, and a line that is not UTF-8 is skipped instead of ending the reader.

A smaller dependency tree

The changelog reports these build sizes before and after:

Build 3.5.0 3.6.0
Default turbomcp 199 crates 175 crates
turbomcp with full 319 crates 295 crates
Default turbomcp-client 235 crates 164 crates
Whole workspace 608 crates 542 crates

Crates present at more than one version went from 43 to 36. The remainder comes from upstream crates (oauth2 on older RustCrypto and rand generations, reqwest 0.13 on tower-http 0.6, dashmap on an older hashbrown). Where a dependency no code used had defined a feature, the feature stays and enables nothing.

What was removed:

  • Unused dependencies. turbomcp-client no longer depends on reqwest, chrono and uuid. turbomcp-macros, a proc-macro crate, dropped axum, tokio, turbomcp-transport, turbomcp-protocol, schemars and serde, which had been compiled a second time for the host. turbomcp-proxy’s rest feature dropped utoipa and utoipa-swagger-ui, and turbomcp-wasm no longer pulls in the whole native turbomcp crate. Twelve other crates lost one or more unused dependencies.
  • One SIMD JSON library. The default simd feature built both simd-json and sonic-rs. The SimdJson serialization format now parses with sonic-rs, and simd-json is gone.
  • One crypto provider. turbomcp-grpc’s default tls feature built tonic’s rustls on ring while reqwest and jsonwebtoken use aws-lc-rs. It now uses tonic/tls-aws-lc, and with turbomcp-dpop’s ring dependency gone, no build compiles ring.
  • No response decompression. The HTTP client no longer builds brotli and gzip decoders, so it sends no Accept-Encoding and servers answer uncompressed. The server’s tower-http no longer builds encoders for a compression layer nothing installed.
  • No rand in a default build. Retry jitter uses fastrand, and generate_secure_token reads the OS random source through getrandom.
  • One tungstenite. The WebSocket server accepted connections through axum’s ws feature, which carries tungstenite 0.29, while the client uses 0.30. The server now performs the RFC 6455 handshake itself over hyper’s upgrade and runs on the same tokio-tungstenite as the client, with the same frame limits.
  • Latest internal majors. OpenTelemetry 0.33 with tracing-opentelemetry 0.34 in turbomcp-telemetry, and nix 0.31 in turbomcp-transport. None of their types appear in a public signature.

Shared line framing

A new turbomcp_transport_traits::codec module (feature codec) provides BoundedLines, the size-limited newline framing that skips an oversized line instead of ending the stream, and its Line frame type. The stdio, TCP and Unix transports each carried an identical private copy and now share this one.

Upgrade notes

[dependencies]
turbomcp = "3.6"

Two changes can break a build, both small:

  1. turbomcp-dpop no longer implements From<ring::error::Unspecified> for DpopError. The crate never called ring, and the dependency existed only for that impl. If you converted a ring error into a DpopError with ?, map it explicitly.
  2. turbomcp-server’s websocket feature no longer enables the WebSocket client transport. A server-only WebSocket build no longer compiles a client. If you relied on the old behaviour, enable turbomcp-transport/websocket yourself. The turbomcp crate’s websocket feature still enables both.
[dependencies]
turbomcp-server = { version = "3.6", features = ["websocket"] }
turbomcp-transport = { version = "3.6", features = ["websocket"] }

Otherwise no public API changed. Behaviour you may notice: TCP and Unix send now applies backpressure instead of silently dropping, HTTP graceful shutdown now ends listening GET streams, and Accept-Encoding is no longer sent by the HTTP client.

Also in this release, internally: the workspace-root tests/ directory is removed (17 files written against v2-era APIs, never compiled by any cargo target since the root manifest is a virtual workspace), and Cargo.lock was refreshed. worker stays on 0.8.6 because 0.8.7 needs Rust 1.91 and the MSRV is 1.89.

Where v3 and v4 stand

v3 is the maintained legacy line, and this release is an example of that work: correctness, security and stability fixes with no feature removals. v4 is in alpha on the main branch and is where breaking work belongs. If you are starting a new production service today, v3 is the supported choice, and we recommend following our v4 coverage as it develops.

Questions about migrating, or support for a v3 deployment? See the TurboMCP product page, the TurboMCP repository, or contact us.

  • TurboMCP
  • Rust
  • MCP
  • Security
  • Transports
  • Dependencies
  • v3

Keep reading

Want this running in your business?

We build and deploy production AI for enterprises. Let's talk about what you're working on.