LangChain

Integrate AgentGuard directly into LangChain at the tool-call layer — no Virtue Gateway required.

Before you start

Review the AgentGuard SDK reference for client setup, result fields, and error handling.

Installation

pip install agentsuite-sdk[langchain]

The [langchain] extra does not include an LLM provider — install the one you need separately (e.g. langchain-openai, langchain-anthropic, langchain-google-genai). The model string follows LangChain's init_chat_model format (e.g. openai:gpt-4o).

Quickstart

1. Guard client

from agentsuite import ActionGuardClient, GuardError

guard = ActionGuardClient(api_key=..., policy_id=...)

2. AgentMiddleware

from langchain.agents import AgentMiddleware, ToolCallRequest
from langchain_core.messages import ToolMessage

class ActionGuardMiddleware(AgentMiddleware):
    tools: list = []

    def __init__(self, guard_client, fail_open: bool = False):
        super().__init__()
        self._guard = guard_client
        self._fail_open = fail_open

    def wrap_tool_call(self, request: ToolCallRequest, handler):
        call = request.tool_call
        tool_name = call.get("name", "unknown")
        result = self._guard.actions.guard_query(
            query=f"Tool: {tool_name}, Args: {call.get('args', {})}",
        )
        if not result.allowed:
            return ToolMessage(..., status="error")
        return handler(request)

    # ... awrap_tool_call — same logic with await handler(request)

3. MCP tools

from langchain_mcp_adapters.client import MultiServerMCPClient

def mcp_connection_config(url: str, mcp_api_key: str | None = None) -> dict:
    conn = {"transport": "streamable_http", "url": url, "timeout": 30}
    if mcp_api_key:
        conn["headers"] = {"X-API-Key": mcp_api_key}
    return {"mcp": conn}

async def get_mcp_tools(mcp_server_url: str, mcp_api_key: str | None = None):
    client = MultiServerMCPClient(mcp_connection_config(mcp_server_url, mcp_api_key))
    return await client.get_tools()

4. Agent and invoke

from langchain.agents import create_agent
from langchain_core.messages import HumanMessage

async def main():
    tools = await get_mcp_tools("http://localhost:3002/mcp")
    agent = create_agent(
        model="openai:gpt-4o",
        tools=tools,
        system_prompt="...",
        middleware=[ActionGuardMiddleware(guard, fail_open=False)],
    )
    result = await agent.ainvoke({"messages": [HumanMessage(content="...")]})

Full runnable example: demo_action_guard_langchain.py

Demo Run

Start the local MCP server (python local_mcp_server.py from the repo root), set env vars, then run:

python examples/demo_action_guard_langchain.py

Environment Variables

Variable	Description
VIRTUE_API_KEY	VirtueAI API key
ACTION_GUARD_POLICY_ID	Policy set ID (agp_...)
OPENAI_API_KEY	OpenAI API key
MCP_SERVER_URL	MCP server URL (default in demos: http://localhost:3002/mcp)
MCP_API_KEY	Optional API key for the MCP server
AGENT_MODEL	Optional model id (default in demo: openai:gpt-4o)