Deprecated documentation

The remote generation alpha has been superseded by remote packages and remote plugins.

For existing users, please see the Migrating from remote generation alpha guide. If you run into issues, contact us on Buf Public Slack.

A common frustration when working with Protobuf is that you need to generate code for each language that you're working with. Many teams implement custom tooling and scripts to solve this problem, but it can be difficult to ensure that every person who works on a given project has all of the code generation tooling set up locally. And if you have Protobuf-based APIs, the consumers of your APIs shouldn't have to deal with code generation.

The Buf Schema Registry solves this problem with remote code generation. With this feature, you can eliminate code generation from your workflows and directly install code generated from Protobuf definitions using standard package managers and build tools. This diagram illustrates how remote generation works:

BSR module
The Buf Schema Registry's remote plugin process

In essence, you can use generation templates to generate code stubs from Buf modules that you've pushed to the BSR. All code generation happens on the BSR itself—not on your laptop, not in a CI/CD environment, only remotely on the BSR.

Remote plugin concepts

Remote plugins in the BSR revolve around a few core concepts:

  • Protobuf plugins generate code from Protobuf definitions
  • Generation templates are named and versioned collections of plugins
  • Remote plugin registries provide interfaces for language-specific tools to install generated SDKs


The BSR uses Protobuf plugins to generate code stubs from Protobuf definitions. Examples of Protobuf plugins include protoc-gen-go and protoc-gen-python.

Creating your own plugins

See the documentation on authoring plugins to see how you can create and upload your own plugins to use as part of code generation.

Plugins belong to an owner and may be public or private. Public plugins are available to anyone, while private plugins are only available to the owner or members of the owning organization. Plugins are often referenced together with their owner's name, for example, library/plugins/protoc-gen-go (or in some contexts just library/protoc-gen-go), is used to reference the protoc-gen-go plugin maintained by Buf.

A plugin has instantiations at different versions. These versions often map directly to the versions of the existing plugin executables. For example, the protoc-gen-go plugin has a version v1.27.1-1 matching the v1.27.1 release of the official Go Protobuf plugin.

Plugin version executables are managed as Docker images. The Docker image is expected to accept a CodeGeneratorRequest in Protobuf binary format on standard in, and respond with a CodeGeneratorResponse in Protobuf binary format on standard out when run. This matches exactly the contract used with existing Protobuf plugins in the ecosystem today, making migration of existing plugins to BSR remote plugins straightforward.

A plugin version is created by pushing a tagged Docker image to the plugins Docker registry repository. For example, assuming the relevant Dockerfile and context was in the current directory, to push a new version v1.1.0 of the plugin protoc-gen-myplugin owned by the user myuser, the user would run

$ docker build -t .

followed by

$ docker push

Pushing plugins to the BSR requires authenticating your Docker CLI using a token:

$ docker login -u myuser --password-stdin

A plugin version can describe runtime library dependencies of its generated assets using Docker labels. All labels are prefixed with build.buf.plugins.runtime_library_versions. followed by the index of the dependency, followed by the attribute being specified. For example, version v1.27.1-1 of the library/protoc-gen-go plugin declares its runtime dependency on the Go module using these labels in its Dockerfile:

LABEL ""=""
LABEL "build.buf.plugins.runtime_library_versions.0.version"="v1.27.1"

You need to give plugins a valid semantic version.

Remote plugins

A feature that you may also find useful is remote plugins.

While remote code generation is geared toward eliminating the need to generate code stubs at all, remote plugins enable you to generate code stubs locally without needing to install plugins locally.


A template defines a collection of plugins and associated configuration to use when generating code stubs from Protobuf. With templates, you can run multiple plugins together, such as protoc-gen-go and protoc-gen-go-grpc, where the output of protoc-gen-go-grpc depends on the output of protoc-gen-go.

Creating your own templates

See the documentation on authoring templates to see how you can upload your own templates to use as part of code generation.

Templates belong to an owner and can be public or private. Public templates are available to anyone, while private templates are available only to the owner or members of the owner's organization.

Buf maintains several official templates:

A template version defines the plugin versions to use. This enables you to keep templates up to date with new versions of plugins in the template. A template version is of the form v[1-9][0-9]*. The template version makes up part of the synthetic version of remotely generated artifacts.

Template management is designed to discourage introducing breaking changes to consumers. This is why plugin parameters are defined on the template itself rather than on a per-version basis.


A remote generation registry is an artifact registry that enables language-specific dependency management tools to install assets remotely generated by the BSR.

The BSR currently offers two registries:

Go module
npm registryJavascript and

Synthetic versions

A synthetic version combines the template and module versions into a semantic version of this form:

Synthetic version syntax

Within this scheme:

  • There's always a v prefix.
  • The major version is always 1.
  • The minor version (3 in the example) corresponds to the template version (without the v prefix). Template versions increase monotonically and have the form v1, v2, v3...
  • The patch version (5 in the example) corresponds to the module, which is identified by a commit sequence ID that's incremented each time a new version of a module is pushed.

The synthetic version v1.2.10, for example, means that the artifact was generated using v2 of the template and using the commit sequence ID 10 for the module.

Where synthetic versions are used

The BSR applies synthetic versions to all remote-generated code artifacts in the remote generation registry. That currently includes Go packages but will be expanded to other languages.

Enforcing semantic versioning

Although we describe synthetic versions as semantic versions, the BSR doesn't enforce semantic versioning. If you make breaking changes to an asset and push that asset to the BSR, the patch version is incremented in spite of the breaking change, which violates semantic versioning.

In order to preserve semver guarantees in your own generated assets, we recommend performing breaking change detection before pushing a new version of a Buf module, potentially as part of your CI/CD pipeline.


Every time you push a Buf module to the BSR, a new commit is created. Each commit has two pieces of information attached to it:

  • A commit name. This is a randomly generated, fixed-size hexadecimal string that's visible in the BSR's UI. Note that the commit name is not a hash of the commit's content.
  • A commit sequence ID. This is a monotonically increasing integer that begins at 1 and is incremented with each new module push. Commit sequence IDs are not visible in the BSR UI.

How we implemented synthetic versions

The challenge with versioning remote-generated code is that unlike versioning schemes that only deal with one artifact, such as a Python library, BSR versions are the product of two logical inputs:

  • The template version
  • The Protobuf module

When implementing our versioning scheme, we surveyed some popular language registries and found that the most common scheme was semantic versioning but without pre-release and build labels. In other words, we found that versions like v1.2.3 were common whereas v1.2.3-alpha.1 were not, and we opted for the former.