Protoc Gen Go Mcp

What is Protoc Gen Go Mcp?

Protoc Gen Go Mcp is a Protocol Buffers compiler plugin that generates Model Context Protocol (MCP) servers for gRPC or ConnectRPC APIs. It auto-generates JSON Schema for tool inputs and is MCP-library-agnostic, supporting both the official Go SDK and mark3labs/mcp-go.

How to use Protoc Gen Go Mcp?

Add the plugin to your buf.gen.yaml, then run buf generate to produce *.pb.mcp.go files. In your Go server code, import the adapter for your chosen MCP library (gosdk or mark3labs), create a server, and register handlers using generated Register*Handler functions. Forward tool calls to gRPC or ConnectRPC clients with generated ForwardTo*Client functions.

Key features of Protoc Gen Go Mcp

• Auto-generates MCP handlers from .proto services
• Outputs JSON Schema for method inputs
• Wire up to gRPC or ConnectRPC servers/clients
• Easy integration with buf build tool
• Runtime LLM provider selection (Standard/OpenAI)
• MCP library agnostic (official SDK and mark3labs)

Use cases of Protoc Gen Go Mcp

• Expose existing gRPC APIs as MCP tools for AI assistants
• Forward MCP tool calls to remote gRPC clients with zero boilerplate
• Namespace multiple service instances (e.g., different databases) via tool name prefixing
• Use with buf to generate MCP servers alongside standard .pb.go files

FAQ from Protoc Gen Go Mcp

Which MCP Go libraries are supported?

Both the official modelcontextprotocol/go-sdk and mark3labs/mcp-go are supported through adapter packages; no regeneration needed to switch.

How are MCP tool names generated?

Tool names default to the full proto method name with dots replaced by underscores. Names longer than 64 characters are truncated using a 10-character hash as prefix, preserving the most specific part.

Can I use custom tool names?

Yes, set the mcp.v1.tool_name method option in your .proto file. The value must match ^[a-z_][a-z0-9_-]{0,63}$; duplicates within a service cause a fallback to the derived name.

How do I migrate from versions before v0.2?

Replace direct mark3labs/mcp-go imports with the adapter package, use WithNamePrefix instead of manual AddTool, and update HandleError return type to *runtime.CallToolResult.

What is the difference between Standard and OpenAI schema modes?

Standard mode supports full JSON Schema (additionalProperties, anyOf, oneOf) and maps as objects. OpenAI mode restricts JSON Schema, converts maps to key-value arrays, and encodes well-known types as JSON strings.