Generating code

Pre-1.0 software

Connect for Rust is pre-1.0. It passes the full conformance suite, but the API surface may still shift within the 0.x series. Join us on Slack if you have questions or feedback.

connect-rust offers two code generation workflows. Both generate the same traits, clients, and message types, so the choice is mostly about how generation fits into your build:

• connectrpc-build runs the generator from build.rs during cargo build. Nothing is checked in, and there are no plugin binaries to install. This is what Getting started uses.
• buf generate runs the generator as a protoc plugin. Generated code lands in your source tree and is committed. Use this when you want the output under review, when you’re generating several languages from one schema, or when you need message types and service stubs in separate module trees.

Generation produces two kinds of output: message types from buffa, and service stubs from connect-rust, which are a trait to implement and a client to call.

Generating from build.rs

Add the build dependency:

[build-dependencies]
connectrpc-build = "0.9"

Then describe your schema in build.rs:

fn main() {
    connectrpc_build::Config::new()
        .files(&["proto/greet/v1/greet.proto", "proto/billing/v1/billing.proto"])
        .includes(&["proto/"])
        .include_file("_connectrpc.rs")
        .compile()
        .unwrap();
}

Message types and service stubs land in one file per proto, written to Cargo’s OUT_DIR and stitched together by the file you named in include_file. Pull the whole tree into your crate with one macro:

pub mod proto {
    connectrpc::include_generated!();
}

Choosing a compiler

connectrpc-build shells out to protoc, which must be on your PATH (or named by the PROTOC environment variable). Two alternatives are available if that doesn’t suit your build.

.use_buf() invokes buf build instead, which is convenient when your schema already has a buf.yaml and remote dependencies. Two things change under it. .includes() is ignored, because buf resolves imports from buf.yaml. And the paths in .files() are used twice: once as buf build --path arguments, resolved against the crate root, and again to match the names buf records in the descriptor set, which are relative to the buf module root. Those two agree only when the module is rooted at the crate root, so put a buf.yaml there:

version: v2
modules:
  - path: .

A module rooted at proto/ instead fails with file_to_generate 'proto/greet/v1/greet.proto' not found in descriptor set.

.descriptor_set(path) reads a precompiled FileDescriptorSet, so neither tool needs to be present at build time. Produce the file once with protoc --descriptor_set_out=... --include_imports.

By default the config emits cargo:rerun-if-changed directives for your proto files, so edits trigger regeneration.

Generating with buf generate

This path uses three plugins: protoc-gen-buffa for message types, protoc-gen-connect-rust for service stubs, and protoc-gen-buffa-packaging to assemble the mod.rs tree for each output directory. The two code generation plugins run per file; only the packaging plugin needs strategy: all.

Installing the plugins

protoc-gen-buffa and protoc-gen-buffa-packaging ship from the buffa repository. For protoc-gen-connect-rust you have three options:

1. Download a pre-built binary from the connect-rust releases page. Linux, macOS, and Windows builds are published, each signed and attested so gh attestation verify or cosign verify-blob can check what you installed.
2. Build from source with cargo install --locked connectrpc-codegen, which puts the binary in $CARGO_HOME/bin.
3. Use the remote plugin on the Buf Schema Registry, buf.build/connectrpc/rust, and skip the local install entirely.

Configuring buf.gen.yaml

version: v2
plugins:
  - local: protoc-gen-buffa
    out: src/generated/buffa
    opt: [views=true, json=true]
  - local: protoc-gen-buffa-packaging
    out: src/generated/buffa
    strategy: all
  - local: protoc-gen-connect-rust
    out: src/generated/connect
    opt: [buffa_module=crate::proto]
  - local: protoc-gen-buffa-packaging
    out: src/generated/connect
    strategy: all
    opt: [filter=services]

To use the remote plugin instead of a local binary, replace the local: protoc-gen-connect-rust entry with remote: buf.build/connectrpc/rust:v0.9.0.

Mount both trees in your crate:

src/lib.rs

#[path = "generated/buffa/mod.rs"]
pub mod proto;
#[path = "generated/connect/mod.rs"]
pub mod connect;

buffa_module=crate::proto tells the service stub generator where you mounted the message types, which is what ties the two trees together. A method whose input is greet.v1.GreetRequest is emitted as crate::proto::greet::v1::GreetRequest: the module root you named, then the Protobuf package as nested modules, then the type. Changing the mount point means regenerating.

The second packaging invocation passes filter=services so the connect tree’s mod.rs only includes files that actually contain service stubs.

buffa_module=X is shorthand for extern_path=.=X, the same option the Buf Schema Registry uses when it generates Cargo SDKs. Any module an extern_path points at has to contain buffa-generated code with views enabled. The service stubs rely on the view types and JSON impls that buffa emits alongside each message. One consequence is worth knowing before you pick this layout: view-body impls are not emitted for types reached through extern_path, so returning a view body needs encodable_impls=all_messages on the crate that owns the type.

Inclusion patterns side by side

The most visible difference between the two workflows is how generated code enters your crate:

// connectrpc-build (build.rs) users:
pub mod proto { connectrpc::include_generated!(); }

// buf generate users:
#[path = "generated/proto/mod.rs"]
pub mod proto;

One file per package

By default, buf generate output is one file per .proto file plus a per-package stitcher. To get a single file per Protobuf package instead, pass file_per_package to both protoc-gen-buffa and protoc-gen-connect-rust. Each plugin then emits one <dotted.pkg>.rs per package with everything inlined. That layout is what Buf Schema Registry Cargo SDKs and tonic-style build integrations expect.

Under this layout you should drop the protoc-gen-buffa-packaging invocations, since there is nothing left for them to wire together. Keep routing each plugin to its own out: directory: the filenames are shared between them and would silently overwrite in a single directory.

connectrpc-build users get the same behavior from Config::file_per_package(true), which inlines service stubs into buffa’s per-package file. The include file picks up the new filename automatically, so nothing else changes.

Generation options

Both workflows expose the same knobs, as plugin opt: entries for buf generate and as builder methods on connectrpc_build::Config.

Plugin option	connectrpc-build	What it does
no_json	.generate_json(false)	Emit message types without serde derives, for proto-only builds
gate_client_feature	.gate_client_feature(true)	Prefix generated clients with #[cfg(feature = "client")] so server-only builds can drop the transport stack
encodable_impls=all_messages	.encodable_impls(..)	Emit view-body impls for every message, not just those used as RPC types
file_per_package	.file_per_package(true)	One output file per Protobuf package
element_memory_limit=N	BUFFA_ELEMENT_MEMORY_LIMIT=N	Raise buffa’s decode budget when a very large descriptor set exceeds it

Proto-only builds

The Connect protocol supports two codecs: binary Protobuf and Protobuf JSON. The JSON codec requires every message type to implement serde::Serialize and Deserialize, which is why the generator derives them by default. A deployment that only ever speaks binary Protobuf can turn JSON off and shed those derives.

It takes two coordinated settings. First, generate without the derives using no_json (or .generate_json(false)). Second, disable the runtime json feature, which relaxes the message-type bounds so serde-free types still satisfy every handler, router, and client signature:

# `default-features = false` is the only way to drop `json`, so it also drops
# the default compression features (`gzip`, `zstd`). Re-list the ones you still
# want.
connectrpc = { version = "0.9", default-features = false, features = ["server", "gzip", "zstd"] }

A proto-only server rejects JSON at content negotiation, before it reads the request body: application/json and application/connect+json come back as HTTP 415 Unsupported Media Type, and the gRPC JSON content types get a gRPC error status. Handler errors stay JSON-encoded, as the Connect protocol requires regardless of the request codec.

Caution

json is an additive, default-on Cargo feature, so it is only off when every crate in your dependency graph that depends on connectrpc disables it. If any other crate pulls in connectrpc with json enabled, feature unification turns it back on for the whole build and your serde-free types stop compiling with a Serialize is not satisfied error that names the trait rather than the feature. Proto-only mode fits a leaf binary or a fully proto-only dependency graph, not one library inside a mixed workspace.

Descriptor sets for reflection

If you need the compiled FileDescriptorSet at runtime, most commonly to feed the connectrpc-reflection crate, chain .emit_descriptor_set("app.fds.bin") before .compile(). The name must be a bare filename with no path separators. The set is written to OUT_DIR with its full transitive import closure, ready to embed:

let bytes: &[u8] = include_bytes!(concat!(env!("OUT_DIR"), "/app.fds.bin"));