human-mcp

Instructions

Implementation Reference

• human_mcp/mcp_server.py:44-61 (handler)

  The primary handler function for the 'human_eye_tool' MCP tool. Decorated with @mcp.tool(), it generates a unique task ID, stores the observation prompt in the database via db_utils, polls asynchronously for human-provided result using wait_for_task_completion, and returns the observation as a dictionary.

  async def human_eye_tool(prompt: str, ctx: Context) -> Dict[str, str]:
      """人間が目で見て状況を説明したり、特定のものを探したりします。"""
      task_id = str(uuid.uuid4())
      instruction = f"👁️ 目を使って観察: {prompt}"
  
      # タスクをデータベースに追加
      db_utils.add_task(task_id, instruction)
  
      # ログ出力
      sys.stderr.write(f"Human task created: {task_id}. Waiting for completion...\n")
  
      # 結果を待機（非同期ポーリング）
      result = await wait_for_task_completion(task_id)
  
      # ログ出力
      sys.stderr.write(f"Human task {task_id} completed.\n")
  
      return {"observation": result}

• human_mcp/tools.py:5-22 (schema)

  Explicit JSON schema definition for the human_eye_tool, detailing input (prompt string) and output (observation string) structures, descriptions, and requirements. Defined in the HUMAN_TOOLS list.

  {
      "name": "human_eye_tool",
      "description": "人間が目で見て状況を説明したり、特定のものを探したりします。",
      "input_schema": {
          "type": "object",
          "properties": {
              "prompt": {"type": "string", "description": "観察するための指示"}
          },
          "required": ["prompt"]
      },
      "output_schema": {
          "type": "object",
          "properties": {
              "observation": {"type": "string", "description": "人間による観察結果"}
          },
          "required": ["observation"]
      }
  },

• human_mcp/mcp_server.py:41-43 (registration)

  The FastMCP server instantiation and @mcp.tool() decorator on the handler function, which registers the human_eye_tool in the MCP server.

  mcp = FastMCP("human-mcp")
  
  @mcp.tool()