MCP 協議

基於 stdio 的 Model Context Protocol v2024-11-05。JSON-RPC 2.0 傳輸,13 個 C 語言原生工具,零拷貝訊息解析。

v2024-11-05MCP Version
stdioTransport
13Tools
JSON-RPC 2.0Protocol

概述

scorpiox code 將 Model Context Protocol(MCP)實作為原生 C 伺服器。MCP 層使用 JSON-RPC 2.0 框架在 stdio 上運行——AI 客戶端通過 stdin/stdout 連接並透過結構化 JSON 訊息呼叫工具。所有 13 個工具都從 sx_tools.c 編譯到二進位檔案中,零執行時依賴。

# MCP server architecture (from sx_tools.c)

stdin  ──▶  JSON-RPC parser  ──▶  tool dispatcher  ──▶  result  ──▶  stdout

# All tools compiled into single static binary
# Source: scorpiox/libsxutil/sx_tools.c
# Zero runtime dependencies — no Node.js, no Python, no JVM

傳輸層

The MCP server uses stdio as its transport — the simplest, most reliable pipe. The AI client spawns the scorpiox binary and communicates via stdin/stdout using newline-delimited JSON-RPC 2.0 messages. No HTTP server, no WebSocket, no port binding.

# Transport: stdio (stdin/stdout)
# Framing: newline-delimited JSON-RPC 2.0
# Each message is a single JSON object terminated by \n

# Client sends request →
{"jsonrpc": "2.0", "id": 1, "method": "tools/list", "params": {}}

# Server responds ←
{"jsonrpc": "2.0", "id": 1, "result": {"tools": [...]}}

初始化握手

Before calling tools, the client must complete the MCP initialization handshake. This establishes the protocol version and declares capabilities.

# Step 1: Client sends initialize request
{"jsonrpc": "2.0", "id": 1, "method": "initialize", "params": {
  "protocolVersion": "2024-11-05",
  "capabilities": {},
  "clientInfo": {"name": "claude-ai", "version": "1.0"}
}}

# Step 2: Server responds with capabilities
{"jsonrpc": "2.0", "id": 1, "result": {
  "protocolVersion": "2024-11-05",
  "capabilities": {"tools": {}},
  "serverInfo": {"name": "scorpiox-code", "version": "1.0.0"}
}}

# Step 3: Client sends initialized notification
{"jsonrpc": "2.0", "method": "notifications/initialized", "params": {}}

JSON-RPC 方法

The MCP server handles 4 JSON-RPC methods. All tool invocations go through tools/call.

方法 方向 Description
initialize client → server Initialize MCP session with protocol version and client capabilities
notifications/initialized client → server Notification sent after successful initialization handshake
tools/list client → server List all tools available from the MCP server
tools/call client → server Call a specific tool with JSON arguments and receive result
# Example: calling the Bash tool
{"jsonrpc": "2.0", "id": 42, "method": "tools/call", "params": {
  "name": "Bash",
  "arguments": {"command": "ls -la /tmp", "timeout": 5000}
}}

# Server response
{"jsonrpc": "2.0", "id": 42, "result": {
  "content": [{"type": "text", "text": "total 48\ndrwxrwxrwt 12 root root ..."}],
  "isError": false
}}

工具目錄 (13)

All tools are implemented in scorpiox/libsxutil/sx_tools.c and compiled into the static binary. Each tool declares a JSON Schema for input validation.

Bash shell

Execute a bash command

參數: command, timeout
必填: command
ReadImage filesystem

Read an image file and display it visually

參數: file_path
必填: file_path
ReadDocs web

Fetch URL content as markdown via headless browser

參數: url, format, wait_ms
必填: url
WebSearch search

Search the web using 11 engines concurrently

參數: query
必填: query
TaskManager workflow

Manage persistent tasks with create/add/complete

參數: action, query, task_id, step_id, text
必填: action, query
PlanMode workflow

Manage plan mode for complex tasks

參數: action, goal, plan_file
必填: action
AskUserQuestion interaction

Ask interactive multiple-choice questions via TUI

參數: questions_json
必填: questions_json
InvokeSkill workflow

Invoke a skill by name for reusable workflows

參數: skill_name, arguments
必填: skill_name
BackgroundBash shell

Execute bash command in background, returns immediately

參數: command
必填: command
ReadBgOutput shell

Read new output from a background task

參數: task_id
必填: task_id
ListBgTasks shell

List all background tasks with status and runtime

參數:
必填:
KillBgTask shell

Kill a running background task by ID

參數: task_id
必填: task_id
SetCallback workflow

Manage scheduled callback messages with repeats

參數: action, message, delay_seconds, repeat_count, callback_id
必填: action

工具模式參考

Complete input_schema for each tool, as declared in the tools/list response.

Bash shell
參數TypeStatusDescription
command string required The bash command to execute
timeout integer optional Timeout in milliseconds (default: 30000)
ReadImage filesystem
參數TypeStatusDescription
file_path string required The absolute path to the image file
ReadDocs web
參數TypeStatusDescription
url string required The URL to fetch
format string optional Output format: text, html, or markdown
wait_ms integer optional Wait time for page load in ms
WebSearch search
參數TypeStatusDescription
query string required The search query
TaskManager workflow
參數TypeStatusDescription
action string required Action: create, add, complete, list, pending, switch
query string required The search query
task_id integer optional Task ID
step_id integer optional Step ID
text string optional Task name or step text
PlanMode workflow
參數TypeStatusDescription
action string required Action: enter, exit, status, json, reset
goal string optional Goal/description
plan_file string optional Path to markdown plan file
AskUserQuestion interaction
參數TypeStatusDescription
questions_json string required JSON string with questions and options
InvokeSkill workflow
參數TypeStatusDescription
skill_name string required The skill name to invoke
arguments string optional Arguments to pass to the skill
BackgroundBash shell
參數TypeStatusDescription
command string required The bash command to execute
ReadBgOutput shell
參數TypeStatusDescription
task_id integer required The task ID returned by BackgroundBash
KillBgTask shell
參數TypeStatusDescription
task_id integer required The task ID to kill
SetCallback workflow
參數TypeStatusDescription
action string required Action: set, list, cancel
message string optional Callback message
delay_seconds integer optional Seconds to wait (1-3600)
repeat_count integer optional Times to fire before cancel
callback_id integer optional Callback slot ID

設定

MCP is enabled by default. The server advertises tools capability during initialization. Configuration is handled via scorpiox-env.txt.

# scorpiox-env.txt — MCP-related settings

TOOLS=true                  # Enable MCP tool serving (default: true)
STREAMING=true              # Stream tool results as they arrive
MODEL=sonnet                # AI model for tool-use conversations

# The MCP server binary path is auto-detected:
# Linux:   ~/.scorpiox/scorpiox-code
# macOS:   ~/.scorpiox/scorpiox-code
# Windows: %USERPROFILE%\.scorpiox\scorpiox-code.exe

# Client config (e.g., Claude Desktop mcp_servers.json):
# {"scorpiox": {"command": "~/.scorpiox/scorpiox-code", "args": []}}