提示词管理MCP服务器

🧩 mcp-prompt-server-go

Your prompt library, served as MCP tools — hot-reloaded, version-controlled, yours.

A single-binary Model Context Protocol server written in Go: every YAML/JSON file in prompts/ becomes a callable MCP tool. 19 ready-made templates + 2 management tools, 1,021 lines of Go, 2 direct dependencies. Edit a file, the server reloads — no restart, no redeploy.

Quickstart · Prompts · Client setup · 中文

Your best prompt is stuck in a chat box

• You wrote a killer code-review prompt once. Now it's buried in a conversation from March.
• You paste the same 500-word instruction into Cursor, then Raycast, then wherever — and they drift apart.
• You tweak a prompt, restart the client, lose your place, repeat.

This server makes prompts files: diffable, versionable, shareable — and callable from any MCP client.

$ echo '{"jsonrpc":"2.0","id":1,"method":"tools/call",
          "params":{"name":"get_prompt_names","arguments":{}}}' | ./bin/mcp-prompt-server

"可用的prompts (19):
- code_review
- gen_html_web_page
- writing_assistant
- ..."

21 tools on first boot: 19 prompt templates from prompts/ + reload_prompts + get_prompt_names.

✨ What it does

Deliberately boring: stdio transport, JSON-RPC, no daemon, no database, no network. ps shows one process; kill is the off switch.

🚀 Quickstart

1. Build (Go 1.21+)

git clone https://github.com/anthonyhann/mcp-prompt-server-go
cd mcp-prompt-server-go && make build

2. Register the binary as a stdio MCP server in your client (full configs below).

3. Call any prompt by name — e.g. code_review with {"code": "..."}.

For AI Agents — paste this to your agent:

Clone https://github.com/anthonyhann/mcp-prompt-server-go, run make build, then register <repo>/bin/mcp-prompt-server as a stdio MCP server in my client config, and verify with the get_prompt_names tool.

🔌 Client setup

Cursor — ~/.cursor/mcp_config.json:

{
  "mcpServers": {
    "prompt-server": {
      "command": "/absolute/path/to/mcp-prompt-server-go/bin/mcp-prompt-server"
    }
  }
}

Windsurf — ~/.codeium/windsurf/mcp_config.json: same shape.

Raycast — "Install Server (MCP)": Name prompt, Command = absolute binary path, no arguments.

Verify any setup:

echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | ./bin/mcp-prompt-server

📚 Built-in prompts

19 templates in 5 families — starting points, meant to be edited in place:

Your own template is a YAML file away:

name: my_prompt
description: What it does
arguments:
  - name: input_text
    description: The input
    required: true
messages:
  - role: user
    content:
      type: text
      text: "Process this: {{input_text}}"

Save → the file watcher reloads → the tool is live.

📁 Repository layout

mcp-prompt-server-go/
├── main.go              # Entry: load prompts → register tools → serve stdio
├── internal/
│   ├── mcp/             # MCP protocol models
│   ├── prompt/          # Template models + manager (fsnotify hot reload, RWMutex)
│   └── server/          # stdio JSON-RPC server
├── prompts/             # 19 built-in YAML templates (edit these!)
├── tools/test_mcp.go    # Manual test harness
├── Makefile             # build / dev / test / package
└── GUIDE.md · UPGRADE.md

⚠️ When NOT to use it

• You want a prompt marketplace. This serves your files; the 19 built-ins are a starter kit, not a feed.
• You need remote access. stdio only — no HTTP/SSE, no auth, no multi-tenancy. One user, one machine.
• Your client only supports the MCP prompts capability. This server exposes prompts as tools (tools/call), which every major client supports — but know the difference.
• You need prompt chaining/workflows. {{var}} substitution is the whole templating engine, on purpose.

🙋 Why it exists

The Node.js original (gdli6177/mcp-prompt-server) proved the idea; I wanted the same thing as one static binary I could drop anywhere without a runtime. So: 1,021 lines of Go, 2 direct deps, make build → done. I run it daily; every built-in prompt here survived real use.

⭐ If it saves you one copy-paste a day, a star tells me to keep the templates coming.

License

MIT © 2026 anthonyhann — protocol: Model Context Protocol. Hat tip to the Node.js originals that inspired this rewrite.