# mcp-go — Build MCP Servers in Go

> mcp-go is a Go SDK for building MCP servers (tools, resources, prompts) and serving them over stdio transports, so agents can call your Go code as tools.

## Install

Merge the JSON below into your `.mcp.json`:

## Quick Use

1. Create a scratch server:
   ```bash
   mkdir mcp-go-demo && cd mcp-go-demo
   go mod init example.com/mcp-go-demo
   go get github.com/mark3labs/mcp-go
   ```
2. Add a minimal server (see README Quickstart) and run:
   ```bash
   go run .
   ```
3. Connect your MCP client to the stdio transport and call the `calculate` tool.

## Intro

mcp-go is a Go SDK for building MCP servers (tools, resources, prompts) and serving them over stdio transports, so agents can call your Go code as tools.

- **Best for:** Go teams who want first-class tool servers (strong typing, middleware, session handling) for agent workflows
- **Works with:** Go 1.x toolchain, stdio MCP clients, server-side tool/resource/prompt patterns
- **Setup time:** 12 minutes

## Practical Notes

- Quickstart includes a calculator tool example you can run in <15 minutes
- The README contains multiple example sections (server/resources/tools/prompts/session management) you can copy incrementally

## A Go-Native Pattern for Tool Design

Treat each MCP tool as a stable “function contract”:

- Keep inputs small and typed (string/number/enum) so the agent can call it reliably.
- Put your real side effects behind a thin validation layer.
- Prefer idempotent operations (safe retries) and provide clear error messages.

Once you have one tool, scale to a toolkit:

1. Add resources for read-only context (like listing projects or configs).
2. Add tools for side effects (create/update/apply).
3. Add prompts to encode standard operating procedures for your team.

That split (resources vs tools vs prompts) makes agent behavior easier to reason about and audit.

### FAQ

**Q: Is mcp-go only for servers?**
A: It’s primarily a server SDK; it helps you expose tools/resources/prompts to MCP clients.

**Q: Do I need HTTP?**
A: No. The README quickstart uses stdio transport, which works with many clients.

**Q: How do I keep tools safe?**
A: Validate inputs, keep side effects minimal, and prefer idempotent operations.

## Source & Thanks

> Source: https://github.com/mark3labs/mcp-go
> License: MIT
> GitHub stars: 8,682 · forks: 826

---

<!-- ZH -->

## 快速使用

1. 创建一个最小示例：
   ```bash
   mkdir mcp-go-demo && cd mcp-go-demo
   go mod init example.com/mcp-go-demo
   go get github.com/mark3labs/mcp-go
   ```
2. 按 README 的 Quickstart 粘贴最小 server 代码，然后运行：
   ```bash
   go run .
   ```
3. 把 MCP 客户端连到 stdio transport，测试调用 `calculate` 工具。

## 简介

mcp-go 是 Go 语言的 MCP SDK：用统一协议把你的 Go 代码暴露为 tools/resources/prompts，并通过 stdio 等传输交给 AI 客户端调用，适合把业务能力封装成可审计的工具层。

- **适合谁：** 想用强类型 Go 写工具服务（tool server），并用于 Agent 工作流的团队
- **可搭配：** Go 工具链、stdio MCP 客户端、server 侧 tools/resources/prompts 模式
- **准备时间：** 12 分钟

## 实战建议

- Quickstart 提供可直接跑通的计算器 tool 示例，通常 <15 分钟即可冒烟
- README 里按主题给了多段示例（server/resources/tools/prompts/session），可逐步拷贝扩展

## 用 Go 写“可控工具”的建议

把每个 MCP tool 当成一个稳定的“函数契约”：

- 输入尽量小且强类型（string/number/enum），让 Agent 更可靠地调用。
- 把真实副作用放在薄薄的一层校验之后（避免脏输入直达系统）。
- 优先做幂等操作（可安全重试），错误信息写清楚（便于 Agent 自修复）。

从一个 tool 扩展为一套工具箱的路径：

1. 先加 resources 提供只读上下文（列项目/配置等）。
2. 再加 tools 做有副作用的动作（create/update/apply）。
3. 用 prompts 固化团队 SOP（让“怎么用”变成可复用模板）。

resources / tools / prompts 三分法会让 Agent 行为更容易审计与推理。

### FAQ

**mcp-go 只做 server 吗？**
答：主要面向 server 侧：把 tools/resources/prompts 暴露给 MCP 客户端。

**必须用 HTTP 吗？**
答：不需要，README quickstart 使用 stdio 传输，很多客户端都支持。

**怎么让工具更安全？**
答：严格校验输入、收敛副作用、尽量幂等并提供清晰错误信息。

## 来源与感谢

> Source: https://github.com/mark3labs/mcp-go
> License: MIT
> GitHub stars: 8,682 · forks: 826


---
Source: https://tokrepo.com/en/workflows/mcp-go-build-mcp-servers-in-go
Author: MCP Hub