LangChain and LangGraph · Immiscible docs

A guarded tool asks Immiscible before it runs and settles after. It keeps its name, description and schema, so it drops into ToolNode, bindTools / bind_tools and the prebuilt agents unchanged.

Tested against @langchain/core 1.2 with @langchain/langgraph 1.4, and langchain-core 1.6 with langgraph 1.2, through a real ToolNode.

#TypeScript

Shell

npm install @immiscible/sdk @langchain/core @langchain/langgraph zod

TypeScript

import { tool } from '@langchain/core/tools';
import { ToolNode } from '@langchain/langgraph/prebuilt';
import { ChatOpenAI } from '@langchain/openai';
import { z } from 'zod';
import { Immiscible } from '@immiscible/sdk';
import { guardLangChainTools } from '@immiscible/sdk/langchain';

const immiscible = new Immiscible().run({ client: 'langchain' });

const buy = tool(async ({ pence, domain }) => placeOrder(pence, domain), {
  name: 'buy', description: 'Buy groceries', schema: z.object({ pence: z.number().int(), domain: z.string() }),
});

const tools = guardLangChainTools([buy, search], {
  client: immiscible,
  mapToAction: ({ name, args }) => name === 'search' ? null
    : Immiscible.paymentAction({ amount: args.pence, currency: 'GBP', merchant: args.domain, provenance: [{ source: 'user' }] }),
});

const model = new ChatOpenAI({ model: 'gpt-5-mini', configuration: immiscible.gateway.openai() }).bindTools(tools);
const toolNode = new ToolNode(tools);

#Python

Shell

python3 -m venv .venv && . .venv/bin/activate
pip install immiscible langchain-core langgraph

Python

from langchain_core.tools import tool
from langgraph.prebuilt import ToolNode
from immiscible import Immiscible
from immiscible.integrations import guard_langchain_tools

immiscible = Immiscible().run(client="langchain")

@tool
def buy(pence: int, domain: str) -> str:
    """Buy groceries from a supermarket."""
    return place_order(pence, domain)

tools = guard_langchain_tools(
    [buy, search],
    client=immiscible,
    map_to_action=lambda call: None if call.name == "search"
        else Immiscible.payment_action(call.args["pence"], "GBP", call.args["domain"], provenance=[{"source": "user"}]),
)
node = ToolNode(tools)                       # in a StateGraph, or create_react_agent(model, tools)

Both invoke and ainvoke are guarded; the async path makes its HTTP calls in a worker thread. When ToolNode hands over a tool call, a refusal comes back as a ToolMessage with status="error" and the tool call’s id, as ToolNode requires. Point a chat model at the gateway with ChatOpenAI(**immiscible.gateway.openai()) (base URL, key and headers) or ChatOpenAI(http_client=...) with the run’s hooks.

You can also guard the function under @tool:

Python

from immiscible.integrations import guarded

@tool
@guarded(to_action, client=immiscible)
def buy(pence: int, domain: str) -> str:
    """Buy groceries from a supermarket."""

#A person in the loop

The default waits for the person inside the tool call, up to ten minutes. In a LangGraph graph with a checkpointer you may prefer to pause the graph instead:

wait: false / wait=False: refuse at once with “waiting for the person to approve”; route on it, and retry the tool call after approval with the same tool call id (it is the idempotency key, so the retry finds the same action, now allowed).
onApprovalRequired / on_approval_required: call LangGraph’s interrupt({ approve: d.approval.url }) there to suspend the graph until the person answers.

Options are the same as for the OpenAI Agents SDK.

Examples: packages/immiscible-js/examples/langchain.mjs, packages/immiscible-py/examples/langchain_example.py. Both run against the fake with --demo.