Skip to content
← Blog

Release

TurboMCP v3.2.0 and v3.3.0: Streamable HTTP Fixes, Valid Tool Schemas, and a Full Dependency Refresh

TurboMCP 3.2.0 fixes Streamable HTTP hangs, CRLF SSE parsing and a reconnect storm. 3.3.0 fixes #[tool] schemas that strict clients rejected and clears six RUSTSEC advisories.

TurboMCP Engineering Team · · 8 min read

TurboMCP 3.2.0 (2026-08-27) and 3.3.0 (2026-09-09) are two minor releases on the 3.x line with different jobs. 3.2.0 makes the Streamable HTTP transport behave correctly under real clients and proxies. 3.3.0 fixes one bug that could take a server’s entire tool catalog offline for strict clients, and refreshes every dependency.

If you run MCP servers in production, the short version is: upgrade to 3.3.0 or later, read the two “Changed” sections below for the small semver breaks, and expect no other API movement.

v3.2.0: Streamable HTTP correctness

Most of 3.2.0 was contributed by @ForrestThump across eight pull requests (#14, #15, #16, #18, #19, #20, #21, #24). The release is about the HTTP client and server behaving the way the Streamable HTTP transport specification says they should.

Fixes

Problem What changed
The client hung waiting for POST-response SSE streams to close send() used to block until the server ended the request’s SSE stream, stalling clients against servers that keep the stream open after answering. It now returns as soon as the correlated response (matching request ID with a result or error) arrives. Remaining stream messages are still consumed and queued.
CRLF-framed SSE streams broke event parsing The client now normalizes \r\n and \r to \n per the SSE specification before splitting events, fixing interop with servers and proxies that emit CRLF.
Missing primer event on POST-response streams POST-initiated streams once again open with an empty-data event carrying a globally unique, stream-encoding event ID ({session}-{stream}-0). That gives clients a Last-Event-ID for resumption, as the 2025-11-25 transport spec recommends.
SSE reconnect storm Backoff was keyed on whether the GET was accepted, so a server that accepted a stream and then closed it was retried with zero delay indefinitely. Backoff is now keyed on how long the stream actually stayed up.
Routine reconnects logged as errors A server closing an idle stream is behaving correctly, and the client reconnecting is doing its job. That is now logged at debug.
Clients that reuse JSON-RPC request ids were locked out See below.

The reconnect fix was measured on the reporter’s deployment. Before the change a server that closed idle streams drew 7,657 connections in two seconds. In a 90-second window the reporter saw 61 SSE error or warn log lines before the change and none after.

Request id reuse

Every transport kept a per-session set of every request id it had seen, and answered a repeat with -32600 "Request ID already used in this session". The spec rule that ids “MUST NOT have been previously used” binds the requestor, whose job is correlating a response with its request. A receiver’s only obligation is to echo the id back. The check bought the server nothing, made interoperability depend on a client’s id allocation strategy, and the set grew for the life of the session. SEP-2567 re-scopes the constraint to ids that are concurrently outstanding.

Concurrent reuse is still logged as a warning, since two responses carrying one id cannot be matched, but it is no longer refused.

Performance

Transport per-message overhead was reduced, with the changes verified by criterion benchmarks (the changelog does not publish figures, so neither do we):

  • HTTP client transport metrics moved from an RwLock to the same lock-free AtomicMetrics the stdio transport uses.
  • Server SSE broadcast payloads are shared as Arc<str> instead of cloning a String per subscriber.
  • SSE frames are encoded with a single exact-size allocation.
  • The stdio transport slices trimmed lines zero-copy from the read buffer.

The benchmarks ship behind a new non-public internal-bench feature. 3.2.0 also re-exports the McpHandler public-method types so downstream code can name them without depending on internal crates, makes just test work on Windows, and consolidates duplicate dependency versions across the workspace.

3.2.0 upgrade notes

3.2.0 is a minor release because two changes are breaking in a strict semver check:

  1. StreamableHttpClientConfig gained sse_healthy_stream_threshold (default 10 seconds), which decides whether a closed SSE stream counts as a failure worth backing off from. Keep it comfortably under your server’s idle timeout. If you construct this config by naming every field, add ..Default::default().
  2. turbomcp_server::transport::http::SessionManager is no longer public unless the non-public internal-bench feature is on. It was internal plumbing exposed unintentionally, and there is no supported replacement.
let config = StreamableHttpClientConfig {
    // ...your existing fields...
    ..Default::default()
};

v3.3.0: Valid tool schemas and a dependency refresh

The $defs bug

Any #[tool] parameter whose type needs a JSON Schema definition (Vec<T>, Option<Vec<T>>, a struct containing a struct) produced a "$ref": "#/$defs/T" next to a $defs block nested inside the property. A root-relative pointer cannot find a definition that lives inside a property. The macro ran schema_for! once per parameter and inserted each root schema whole under properties.<name>, so each property also carried its own $schema and a title holding the Rust type name.

Validating clients reject the whole request over one such tool. llama.cpp’s llama-server answers HTTP 400 with Error resolving ref #/$defs/T: $defs not in {...}, which takes the server’s entire catalog offline for that host. Non-validating hosts lose grammar-constrained argument generation for the affected tool. It was reported with a repro in Epistates/turbovault#51.

All parameters are now rendered through one shared SchemaGenerator, so a tool schema is a single document:

  • Properties keep the parameter type inlined, as before.
  • Definitions land in one root $defs.
  • $schema and title no longer appear inside properties.
  • Two parameter types that share a name get distinct definitions instead of one overwriting the other.
  • A self-referential type points at #/$defs/T rather than "$ref": "#", which resolved to the tool schema instead of the type.

A tool like the following now produces a schema that strict validators accept:

use turbomcp::prelude::*;

#[derive(serde::Deserialize, schemars::JsonSchema)]
struct Item {
    sku: String,
    quantity: u32,
}

#[derive(Clone)]
struct Orders;

#[server(name = "orders", version = "1.0.0")]
impl Orders {
    #[tool("Place an order")]
    async fn place(&self, items: Vec<Item>) -> McpResult<String> {
        Ok(format!("{} line items", items.len()))
    }
}

TurboMCP 4.x builds one args struct per tool and was never affected.

Security advisories

3.3.0 clears six RUSTSEC advisories and one yanked crate from Cargo.lock:

Crate Change Advisory
h2 0.4.14 to 0.4.19 RUSTSEC-2026-0258, unbounded empty DATA frames. Reachable on every HTTP/2 path through hyper, axum, reqwest and tonic.
rkyv 0.8.16 to 0.8.18 RUSTSEC-2026-0233, 0234, 0235, archive validation. Reached from turbomcp-core.
crossbeam-epoch 0.9.18 to 0.9.21 RUSTSEC-2026-0204
quinn-proto 0.11.14 to 0.11.17 RUSTSEC-2026-0185. In the lockfile through reqwest’s optional HTTP/3 support, which no TurboMCP feature enables.

The four unsound warnings (anyhow, event-listener, memmap2, scc) resolve with the same refresh. cargo audit and cargo deny check are both clean. The yanked crate was metrics 0.24.5, now 0.24.6.

Why 3.3.0 is a minor release

Every direct dependency moved to its current major. Three of those majors appear in public signatures, so a crate that passes those libraries’ types into TurboMCP has to move to the same major. cargo semver-checks reports nothing in TurboMCP’s own surface.

Dependency Move Where it appears Fix if you also depend on it
jsonwebtoken 10 to 11 JwtValidationResult::algorithm and JwtValidator::with_algorithms (turbomcp-auth), the helpers module of turbomcp-dpop, ProxyAuthConfig::with_algorithm (turbomcp-proxy) Bump to 11
tokio-tungstenite 0.29 to 0.30 WebSocketBidirectionalTransport::send_raw_message (turbomcp-websocket) takes a tungstenite::Message Bump to 0.30
tower-http 0.6 to 0.7 turbomcp_proxy::runtime::build_cors_layer returns a CorsLayer Bump to 0.7

Other direct dependencies moved too, including OpenTelemetry 0.31 to 0.32 (no OpenTelemetry type is in a public signature), syn 2 to 3, p256 0.13 to 0.14, sha2 0.10 to 0.11 and rand 0.9 to 0.10. Three crates needed source changes: turbomcp-macros for syn 3, turbomcp-dpop for the p256 0.14 APIs, and turbomcp-cli for comfy-table 8. The MSRV stays at 1.89.0.

Where an older major remains in Cargo.lock, it arrives through a third party (axum 0.8 pins tokio-tungstenite 0.29, reqwest pins tower-http 0.6, yubihsm pins p256 0.13) and cannot be selected from this side.

Upgrading

[dependencies]
turbomcp = "3.3"
  1. If you build StreamableHttpClientConfig with every field named, add ..Default::default().
  2. If you also depend on jsonwebtoken, tokio-tungstenite or tower-http and pass their types into TurboMCP, bump them to 11, 0.30 and 0.7.
  3. If you name turbomcp_server::transport::http::SessionManager, remove the dependency on it.
  4. Re-run your tool listing against a strict client. Servers with Vec<T> or nested-struct parameters should now list cleanly.

For the full text, see the TurboMCP repository, the TurboMCP product page, or get in touch.

  • TurboMCP
  • Rust
  • MCP
  • Streamable HTTP
  • JSON Schema
  • Security
  • 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.