docs: README에 Gateway Tools (search_tools + call_tool) 사용 예제 추가

LangChain 섹션 전면 개편:
- Gateway Tools: 62 tools → 2 meta-tools, 92% 토큰 절감 예제
- Auto-filter Agent: 매 턴 자동 필터링 예제
- 패턴 비교표 (Gateway vs Auto-filter vs Manual)
- At a Glance에 LangChain Gateway 행 추가
- 한국어/중국어/일본어 README 동시 업데이트
- xgen-workflow 시뮬레이션 e2e 테스트 + 토큰 절감 검증 테스트

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: SonAIengine

README-ja.md

@@ -373,12 +373,44 @@ from graph_tool_call.middleware import patch_anthropic
 patch_anthropic(client, graph=tg, top_k=5)
 ```
 
-### LangChain統合
+### LangChain / LangGraph統合
 
 ```bash
-pip install graph-tool-call[langchain]
+pip install graph-tool-call[langchain] langgraph
 ```
 
+#### Gateway Tools（大規模ツールセット推奨）
+
+50~500+個のツールを**2個のmeta-tool**（`search_tools` + `call_tool`）に変換：
+
+```python
+from graph_tool_call.langchain import create_gateway_tools
+
+# 62個のツールを2個のgateway meta-toolに変換
+gateway = create_gateway_tools(all_tools, top_k=10)
+
+# 2個のツールだけagentに渡す
+agent = create_react_agent(model=llm, tools=gateway)
+```
+
+| | 全ツールバインド | Gateway（2個） |
+|---|:---:|:---:|
+| **62 tools** | ~6,090 tokens/turn | ~475 tokens/turn |
+| **トークン削減** | — | **92%** |
+
+#### 自動フィルタリングAgent
+
+毎ターン自動的に関連ツールのみLLMにバインド：
+
+```python
+from graph_tool_call.langchain import create_agent
+
+agent = create_agent(llm, tools=all_200_tools, top_k=5)
+```
+
+<details>
+<summary>LangChain Retriever（Document返却）</summary>
+
 ```python
 from graph_tool_call import ToolGraph
 from graph_tool_call.langchain import GraphToolRetriever
@@ -393,6 +425,8 @@ for doc in docs:
     print(doc.metadata["tags"])   # ["order"]
 ```
 
+</details>
+
 ---
 
 ## ベンチマーク

README-ko.md

@@ -468,19 +468,52 @@ from graph_tool_call.middleware import patch_anthropic
 patch_anthropic(client, graph=tg, top_k=5)
 ```
 
-### LangChain 연동
+### LangChain / LangGraph 연동
 
 ```bash
-pip install graph-tool-call[langchain]
+pip install graph-tool-call[langchain] langgraph
+```
+
+#### Gateway Tools (대규모 tool set 추천)
+
+50~500+ tool을 **2개 meta-tool**로 변환. LLM이 검색 → 실행하는 구조:
+
+```python
+from graph_tool_call.langchain import create_gateway_tools
+
+# Slack, GitHub, Jira, MS365 등 다양한 소스의 tool 62개
+gateway = create_gateway_tools(all_tools, top_k=10)
+# → [search_tools, call_tool] 2개만 agent에 전달
+
+agent = create_react_agent(model=llm, tools=gateway)
+result = agent.invoke({"messages": [("user", "PROJ-123 이슈를 Done으로 변경해줘")]})
+```
+
+| | 전체 tool 바인딩 | Gateway (2개) |
+|---|:---:|:---:|
+| **62 tools** | ~6,090 tokens/turn | ~475 tokens/turn |
+| **토큰 절감** | — | **92%** |
+
+#### 자동 필터링 Agent
+
+매 턴 자동으로 관련 tool만 LLM에 바인딩:
+
+```python
+from graph_tool_call.langchain import create_agent
+
+agent = create_agent(llm, tools=all_200_tools, top_k=5)
+# 매 턴 유저 메시지 기반으로 ~5개 tool만 노출
 ```
 
+<details>
+<summary>LangChain Retriever (Document 반환)</summary>
+
 ```python
 from graph_tool_call import ToolGraph
 from graph_tool_call.langchain import GraphToolRetriever
 
 tg = ToolGraph.from_url("https://api.example.com/openapi.json")
 
-# LangChain retriever — 모든 chain/agent와 호환
 retriever = GraphToolRetriever(tool_graph=tg, top_k=5)
 docs = retriever.invoke("cancel an order")
 
@@ -489,6 +522,8 @@ for doc in docs:
     print(doc.metadata["tags"])   # ["order"]
 ```
 
+</details>
+
 ---
 
 ## 벤치마크

README-zh_CN.md

@@ -372,12 +372,44 @@ from graph_tool_call.middleware import patch_anthropic
 patch_anthropic(client, graph=tg, top_k=5)
 ```
 
-### LangChain 集成
+### LangChain / LangGraph 集成
 
 ```bash
-pip install graph-tool-call[langchain]
+pip install graph-tool-call[langchain] langgraph
 ```
 
+#### Gateway Tools（推荐用于大规模工具集）
+
+将 50~500+ 个工具转换为 **2个 meta-tool**（`search_tools` + `call_tool`）：
+
+```python
+from graph_tool_call.langchain import create_gateway_tools
+
+# 将 62 个工具转为 2 个 gateway meta-tool
+gateway = create_gateway_tools(all_tools, top_k=10)
+
+# 只需 2 个工具传入 agent
+agent = create_react_agent(model=llm, tools=gateway)
+```
+
+| | 绑定全部工具 | Gateway（2个） |
+|---|:---:|:---:|
+| **62 tools** | ~6,090 tokens/turn | ~475 tokens/turn |
+| **Token 减少** | — | **92%** |
+
+#### 自动过滤 Agent
+
+每轮自动只绑定相关工具到 LLM：
+
+```python
+from graph_tool_call.langchain import create_agent
+
+agent = create_agent(llm, tools=all_200_tools, top_k=5)
+```
+
+<details>
+<summary>LangChain Retriever（返回 Document）</summary>
+
 ```python
 from graph_tool_call import ToolGraph
 from graph_tool_call.langchain import GraphToolRetriever
@@ -392,6 +424,8 @@ for doc in docs:
     print(doc.metadata["tags"])   # ["order"]
 ```
 
+</details>
+
 ---
 
 ## 基准测试

README.md

@@ -53,6 +53,7 @@ LLM agents need tools. But as tool count grows, two things break:
 | **Zero dependencies** | Core runs on Python stdlib only — add extras as needed |
 | **Any tool source** | Auto-ingest from OpenAPI / Swagger / MCP / Python functions |
 | **History-aware** | Previously called tools are demoted; next-step tools are boosted |
+| **LangChain Gateway** | 62 tools → 2 meta-tools, **92% token reduction** per turn |
 | **MCP Proxy** | 172 tools across servers → 3 meta-tools, saving ~1,200 tokens/turn |
 
 ---
@@ -524,7 +525,53 @@ toolkit.graph.enable_embedding("ollama/qwen3-embedding:0.6b")
 pip install graph-tool-call[langchain] langgraph
 ```
 
-**Drop-in agent** — pass all your tools, graph-tool-call **automatically filters per turn**:
+Three integration patterns — pick the one that fits your architecture:
+
+#### Gateway Tools (recommended for large tool sets)
+
+Convert 50~500+ tools into **2 meta-tools** (`search_tools` + `call_tool`).
+The LLM searches first, then calls — no tool definitions bloat in context.
+
+```python
+from graph_tool_call.langchain import create_gateway_tools
+
+# 62 tools from Slack, GitHub, Jira, MS365, custom APIs...
+all_tools = slack_tools + github_tools + jira_tools + ms365_tools + api_tools
+
+# Convert to 2 gateway meta-tools
+gateway = create_gateway_tools(all_tools, top_k=10)
+# → [search_tools, call_tool]
+
+# Use with any LangChain agent — only 2 tools in context
+agent = create_react_agent(model=llm, tools=gateway)
+result = agent.invoke({"messages": [("user", "PROJ-123 이슈를 Done으로 변경해줘")]})
+```
+
+**How it works** — the LLM drives the search:
+
+```text
+User: "Cancel order #500"
+  ↓
+LLM calls search_tools(query="cancel order")
+  → returns: cancel_order, get_order, process_refund (with parameter info)
+  ↓
+LLM calls call_tool(tool_name="cancel_order", arguments={"order_id": 500})
+  → returns: {"order_id": 500, "status": "cancelled"}
+  ↓
+LLM: "Order #500 has been cancelled."
+```
+
+| | All tools bound | Gateway (2 tools) |
+|---|:---:|:---:|
+| **62 tools** | ~6,090 tokens/turn | ~475 tokens/turn |
+| **Token reduction** | — | **92%** |
+| **Accuracy** (qwen3.5:4b) | — | 70% (100% with GPT-4o) |
+
+> Works with **any existing LangChain agent setup**. Just replace `tools=all_tools` with `tools=create_gateway_tools(all_tools)`.
+
+#### Auto-filtering Agent (transparent per-turn filtering)
+
+The agent automatically filters tools each turn — the LLM never sees the full list:
 
 ```python
 from graph_tool_call.langchain import create_agent
@@ -537,11 +584,16 @@ result = agent.invoke({"messages": [("user", "cancel my order")]})
 # Turn 2: LLM sees [next relevant tools based on conversation]
 ```
 
-Each turn, the latest user message is used to retrieve relevant tools via ToolGraph,
-and only those are bound to the model — **saving tokens automatically**.
+#### Which to use?
+
+| Pattern | Best for | How it works |
+|---------|----------|--------------|
+| **Gateway** | 50+ tools, existing agents | LLM explicitly searches → calls |
+| **Auto-filter** | New agents, simple setup | Transparent per-turn tool swap |
+| **Manual** | Full control | You call `filter_tools()` yourself |
 
 <details>
-<summary>Manual filtering (more control)</summary>
+<summary>Manual filtering</summary>
 
 ```python
 from graph_tool_call import filter_tools