diff --git a/CLAUDE.md b/CLAUDE.md
index 5e3fa49..3710241 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -440,11 +440,19 @@ uv run python cli.py serve --port 8080
 
 ## HTTP 服务 (server.py，可选)
 
+纯推理服务，不依赖 torch / 训练代码，仅需 onnxruntime + FastAPI。
+
 ```python
-# FastAPI 服务，提供 REST API
-# POST /solve  - 上传图片，返回识别结果
-# 请求: multipart/form-data，字段名 image
-# 响应: {"type": "normal", "result": "A3B8", "confidence": 0.95, "time_ms": 45}
+# POST /solve         - JSON base64 图片识别
+#   请求: {"image": "<base64>", "type": "normal"}   (type 可选)
+#   响应: {"type": "normal", "result": "A3B8", "raw": "A3B8", "time_ms": 12.3}
+#
+# POST /solve/upload  - multipart 文件上传识别
+#   请求: multipart/form-data, 字段名 image, 可选 query param type
+#   响应: 同上
+#
+# GET  /health        - 健康检查
+#   响应: {"status": "ok", "models_loaded": true}
 ```
 
 ## 关键约束和注意事项
diff --git a/server.py b/server.py
index 03b5270..f2e57f0 100644
--- a/server.py
+++ b/server.py
@@ -1,21 +1,30 @@
 """
-FastAPI HTTP 推理服务
+FastAPI HTTP 推理服务 (纯推理，不依赖 torch/训练代码)
 
-提供 REST API 识别验证码:
-    POST /solve  - 上传图片，返回识别结果
-    GET  /health - 健康检查
+仅依赖: fastapi, uvicorn, python-multipart, onnxruntime, pillow, numpy
 
-启动方式:
-    uv run python cli.py serve --port 8080
+API:
+    POST /solve         JSON base64 图片识别
+    POST /solve/upload  multipart 文件上传识别
+    GET  /health        健康检查
 
-请求示例:
-    curl -X POST http://localhost:8080/solve -F "image=@captcha.png"
+启动:
+    uv sync --extra server
+    python cli.py serve --port 8080
 
-响应示例:
-    {"type": "normal", "result": "A3B8", "confidence": 0.95, "time_ms": 45}
+请求示例 (base64):
+    curl -X POST http://localhost:8080/solve \
+      -H "Content-Type: application/json" \
+      -d '{"image": "<base64>", "type": "normal"}'
+
+请求示例 (文件上传):
+    curl -X POST http://localhost:8080/solve/upload -F "image=@captcha.png"
+
+响应:
+    {"type": "normal", "result": "A3B8", "raw": "A3B8", "time_ms": 12.3}
 """
 
-from __future__ import annotations
+import base64
 
 
 def create_app():
@@ -23,10 +32,12 @@ def create_app():
     创建 FastAPI 应用实例（工厂函数）。
 
     cli.py 的 cmd_serve 依赖此签名。
-    需要安装 server 可选依赖: uv sync --extra server
     """
+    from typing import Optional
+
     from fastapi import FastAPI, File, Query, UploadFile
     from fastapi.responses import JSONResponse
+    from pydantic import BaseModel
 
     from inference.pipeline import CaptchaPipeline
 
@@ -36,47 +47,38 @@ def create_app():
         version="0.1.0",
     )
 
-    pipeline: CaptchaPipeline | None = None
+    pipeline: Optional[CaptchaPipeline] = None
 
+    # ---- 启动时加载模型 ----
     @app.on_event("startup")
     def _load_models():
         nonlocal pipeline
         try:
             pipeline = CaptchaPipeline()
         except FileNotFoundError:
-            # 模型未导出时允许启动，但 /solve 会返回 503
             pipeline = None
 
-    @app.get("/health")
-    def health():
-        models_loaded = pipeline is not None
-        return {"status": "ok" if models_loaded else "no_models", "models_loaded": models_loaded}
+    # ---- 请求体定义 ----
+    class SolveRequest(BaseModel):
+        image: str  # base64 编码的图片
+        type: Optional[str] = None  # 指定类型可跳过分类
 
-    @app.post("/solve")
-    async def solve(
-        image: UploadFile = File(...),
-        type: str | None = Query(None, description="指定类型跳过分类"),
-    ):
+    # ---- 通用推理逻辑 ----
+    def _solve(image_bytes: bytes, captcha_type: Optional[str]) -> dict:
         if pipeline is None:
             return JSONResponse(
                 status_code=503,
                 content={"error": "模型未加载，请先训练并导出 ONNX 模型"},
             )
-
-        data = await image.read()
-        if not data:
-            return JSONResponse(
-                status_code=400,
-                content={"error": "空文件"},
-            )
+        if not image_bytes:
+            return JSONResponse(status_code=400, content={"error": "空图片"})
 
         try:
-            result = pipeline.solve(data, captcha_type=type)
-        except RuntimeError as e:
-            return JSONResponse(
-                status_code=400,
-                content={"error": str(e)},
-            )
+            result = pipeline.solve(image_bytes, captcha_type=captcha_type)
+        except (RuntimeError, TypeError) as e:
+            return JSONResponse(status_code=400, content={"error": str(e)})
+        except Exception as e:
+            return JSONResponse(status_code=400, content={"error": f"图片解析失败: {e}"})
 
         return {
             "type": result["type"],
@@ -85,4 +87,34 @@ def create_app():
             "time_ms": result["time_ms"],
         }
 
+    # ---- 路由 ----
+    @app.get("/health")
+    def health():
+        models_loaded = pipeline is not None
+        return {
+            "status": "ok" if models_loaded else "no_models",
+            "models_loaded": models_loaded,
+        }
+
+    @app.post("/solve")
+    async def solve_base64(req: SolveRequest):
+        """JSON 请求，图片通过 base64 传输。"""
+        try:
+            image_bytes = base64.b64decode(req.image)
+        except Exception:
+            return JSONResponse(
+                status_code=400,
+                content={"error": "base64 解码失败"},
+            )
+        return _solve(image_bytes, req.type)
+
+    @app.post("/solve/upload")
+    async def solve_upload(
+        image: UploadFile = File(...),
+        type: Optional[str] = Query(None, description="指定类型跳过分类"),
+    ):
+        """multipart 文件上传。"""
+        data = await image.read()
+        return _solve(data, type)
+
     return app