LangChain

Use BelgieMiddleware to add sandboxed run_code and session management to a LangChain agent. The middleware supports both synchronous and asynchronous LangChain agent execution.

Install

uv add "belgie[langchain]"

Configure the model provider separately using the LangChain documentation.

Add the middleware

agent.py

from langchain.agents import create_agent

from belgie.langchain import BelgieMiddleware

agent = create_agent(
    model="openai:gpt-5",
    tools=[],
    middleware=[BelgieMiddleware()],
    system_prompt=(
        "Use run_code when JavaScript or TypeScript is useful. "
        "Return the value from the exported function."
    ),
)

result = agent.invoke(
    {
        "messages": [
            ("user", "Use TypeScript to convert 'hello-world' to camelCase."),
        ],
    },
)
print(result["messages"][-1].content)

The middleware creates a session before the agent starts, adds Belgie tools to the model request, and closes the session after the agent finishes. The default session denies network access and limits script reads to its workspace.

Configure the middleware

BelgieMiddleware accepts the shared options described in AI agents:

from belgie.langchain import BelgieMiddleware

middleware = BelgieMiddleware(
    max_retries=2,
    timeout=30,
    instructions="Use fetch for HTTP APIs and return JSON-serializable values.",
)

When a Belgie runtime error or timeout occurs, the middleware returns an error ToolMessage for the Belgie tool. Other tools continue through LangChain's normal middleware chain.

Configure permissions

Allow only the hosts required by the agent through RuntimeOptions:

from belgie import RuntimeOptions, RuntimePermissions
from belgie.langchain import BelgieMiddleware

runtime_options = RuntimeOptions(
    permissions=RuntimePermissions(allow_net=["api.example.com"]),
)

middleware = BelgieMiddleware(runtime_options=runtime_options)

Use deferred loading

from belgie.langchain import BelgieMiddleware

middleware = BelgieMiddleware(
    defer_loading=True,
    capability_id="belgie-js",
)

The model first receives load_belgie and can request the full capability when it needs it. Keep the identifier stable across runs when the application uses multiple deferred capabilities.

Async agents and rendering

Use agent.ainvoke(...) for an asynchronous run. Keep the agent invocation active until the middleware has closed its session. TSX scripts can return npm:@belgie/render in the same way as the Pydantic AI integration. The renderer is a separate host-side pass; it does not grant the model-visible script access to host system paths or FFI.

See @belgie/render for widget constraints.