MCP Adapter

精确调用链

HTTP POST /mcp/<secret>
  ↓
JSON-RPC method
  ├─ initialize
  ├─ tools/list
  │    ↓
  │  deps.listTools()
  │    ↓
  │  validate internal identity
  │    ↓
  │  cache Map<name, Tool>
  │    ↓
  │  project inputSchema + annotations
  │
  └─ tools/call
       ↓
     Map lookup
       ↓
     deps.callTool({ identity, args, toolCallId })
       ↓
     mcpResult()
       ↓
     JSON-RPC result

Symbol 边界

Symbol	负责什么	不负责什么
createRoutedMcpBridge()	transport/protocol adaptation	Agent Loop
isReadOnly()	annotation heuristic	authorization
tools Map	discovery→call identity continuity	durable registry
tools/list	schema projection	tool execution
tools/call	lookup + internal dispatch	semantic task success
mcpResult()	result shape normalization	business reconciliation
deps.callTool()	internal executor injection point	MCP protocol semantics

源码还限制 request body 大小、绑定 localhost、用随机 secret path。这些属于 bridge transport hardening，不等于 MCP 的全部安全模型。

本地定位

rg -n "createRoutedMcpBridge|isReadOnly|mcpResult|notifications/initialized|tools/list|tools/call|callTool" source/node-agent-coordinator/routed-mcp-bridge.ts

固定源码

• routed-mcp-bridge.ts

关联知识文章

• 不可信输入怎样限制权限：Tool Result、Permission 与 Capability Security — Result Adapter 只转换数据形状，不授予 capability；mcpResult()
• 外部能力适配：MCP 与 Workspace / Sandbox — MCP 只做 Internal Tool ↔ Protocol 的转换；createRoutedMcpBridge() → tools/list branch → tools/call branch
• 模型计划如何变成真实执行：ToolExecutor、Planner 与 GraphCompiler — 内部 Tool Executor 不应依赖 MCP request 类型；tools/call branch → deps.callTool()
• MCP 如何发现能力：Self-describing Request、Discovery 与 Tools — MCP Discovery 如何投影到 Harness Tool Registry；tools/list branch → tools Map
• MCP 如何接入现有 Harness：Gateway、Conformance 与 Adapter — MCP Bridge 不拥有业务执行，只做协议适配；createRoutedMcpBridge() → tools/list branch → tools/call branch → mcpResult()