[usage] add self-host detailed README

README.md

@@ -166,6 +166,22 @@ For specialized deployment scenarios, see the detailed guides:
 - **[Apple Silicon with mlx-vlm](examples/mlx-deploy/README.md)** - Optimized for Apple Silicon Macs
 - **[Ollama Deployment](examples/ollama-deploy/README.md)** - Simple local deployment with Ollama
 
+#### Option 4: SDK Server + Client (GPU-less Client)
+
+Deploy the SDK Server on a GPU machine, then use any machine as a client — no GPU needed on the client side. The client connects via the MaaS-compatible protocol, pointing `api_url` at your self-hosted server.
+
+```yaml
+# Client config.yaml
+pipeline:
+  maas:
+    enabled: true
+    api_url: http://<SERVER_IP>:5002/glmocr/parse
+    api_key: any-string    # self-hosted server does not validate keys
+    verify_ssl: false
+```
+
+See the full guide: **[Self-hosted SDK Server + Client](examples/self-host/README.md)**
+
 #### Update Configuration
 
 After launching the service, configure `config.yaml`:

README_zh.md

@@ -164,6 +164,22 @@ sglang serve --model zai-org/GLM-OCR --port 8080 --speculative-algorithm NEXTN -
 - **[Apple Silicon 使用 mlx-vlm](examples/mlx-deploy/README.md)** - 针对 Apple Silicon Mac 优化
 - **[Ollama 部署](examples/ollama-deploy/README.md)** - 使用 Ollama 进行简单的本地部署
 
+#### 方式 4: SDK Server + Client（Client 无需 GPU）
+
+在 GPU 机器上部署 SDK Server，其他机器作为 Client 通过网络调用——Client 端无需 GPU。Client 通过 MaaS 兼容协议连接，将 `api_url` 指向自建 Server 即可。
+
+```yaml
+# Client 端 config.yaml
+pipeline:
+  maas:
+    enabled: true
+    api_url: http://<SERVER_IP>:5002/glmocr/parse
+    api_key: any-string    # 自建 Server 不校验 API key
+    verify_ssl: false
+```
+
+完整指南见：**[自部署 SDK Server + Client](examples/self-host/README.md)**
+
 ##### 更新配置
 
 启动服务后，配置 `config.yaml`：

examples/self-host/README.md

@@ -0,0 +1,103 @@
+# Self-hosted SDK Server + Client Mode
+
+The GLM-OCR SDK supports a split **Server / Client** deployment: run the SDK Server on a GPU machine, and connect from any other machine over HTTP — no GPU required on the client side.
+
+## Architecture
+
+```
+┌──────────────────────┐       HTTP        ┌──────────────────────────────┐
+│  Client (no GPU)     │ ────────────────→ │  Server (GPU machine)        │
+│                      │  POST /glmocr/parse│                              │
+│  glmocr CLI / Python │ ←──────────────── │  python -m glmocr.server     │
+│                      │  JSON response     │  (layout + OCR pipeline)    │
+└──────────────────────┘                    └──────────────────────────────┘
+```
+
+The Server runs the full OCR pipeline (layout detection + parallel OCR). The Client calls it over HTTP with zero local computation.
+
+## Server Side
+
+On the GPU machine, start the Server:
+
+```bash
+# 1. Install (includes selfhosted pipeline + server)
+pip install "glmocr[selfhosted,server]"
+
+# 2. Configure self-hosted mode and point to your local vLLM / SGLang
+#    In config.yaml: set pipeline.maas.enabled: false and configure ocr_api
+
+# 3. Start the server
+python -m glmocr.server --config config.yaml
+```
+
+The server listens on `0.0.0.0:5002` by default, with the API endpoint at `/glmocr/parse`.
+
+## Client Side
+
+On any machine (no GPU needed), point the SDK's MaaS client at your self-hosted server:
+
+```bash
+pip install glmocr
+```
+
+Edit `config.yaml`:
+
+```yaml
+pipeline:
+  maas:
+    enabled: true
+    api_url: http://<SERVER_IP>:<SERVER_PORT>/glmocr/parse
+    api_key: any-string        # Self-hosted server does not validate API keys
+    verify_ssl: false          # Internal networks typically lack HTTPS
+```
+
+Then use the CLI or Python API:
+
+```bash
+# CLI
+glmocr parse document.png --config config.yaml
+
+# Or set via environment variable
+export ZHIPU_API_KEY=any-string
+glmocr parse document.png
+```
+
+```python
+# Python API
+from glmocr import GlmOcr
+
+with GlmOcr(
+    mode="maas",
+    api_url="http://<SERVER_IP>:5002/glmocr/parse",
+    api_key="any-string",
+) as parser:
+    result = parser.parse("document.png")
+    print(result.markdown_result)
+```
+
+## Protocol Details
+
+The server accepts both input formats:
+
+| Input format | Example |
+|---|---|
+| SDK native | `{"images": ["url1", "url2"]}` |
+| MaaS compatible | `{"file": "url", "model": "glm-ocr"}` |
+
+The server response includes both SDK and MaaS field sets:
+
+```json
+{
+  "json_result": [...],
+  "markdown_result": "...",
+  "layout_details": [...],
+  "md_results": "...",
+  "data_info": {"pages": []},
+  "usage": {},
+  "model": "glm-ocr",
+  "id": "chatcmpl-...",
+  "created": 1709234567
+}
+```
+
+This means the client can use the SDK CLI / Python API or send raw HTTP requests — both will parse the response correctly.