langchain-mnki

LangChain / LangGraph tools guarded by mnki (Agent Trust): every tool call is verified against the agent's delegated authority, with an evidence list, before it runs.

pip install langchain-mnki

import os
from mnki import AgentTrustClient, Guard
from langchain_mnki import guard_tools
from langgraph.prebuilt import create_react_agent

g = Guard(AgentTrustClient("https://mnki.com", api_key=os.environ["MNKI_API_KEY"]), agent="invoice-agent")
graph = create_react_agent(model, guard_tools(g, [refund, lookup]))

• Tools stay BaseTool instances (a GuardedTool subclass), so ToolNode, bind_tools and isinstance checks work.
• A refused call returns {"error": "denied", "reasons": [...], "evidence": [...]} to the model; on_refused="throw" raises.
• Calls that policy escalates wait for a human approval (phone app) and continue when approved.
• No account needed to try it: Guard(world=demo_world(), agent="invoice-agent") runs the same verifier offline.

Pinned versions and the support level are in the compatibility matrix; every integration path is on one page at https://mnki.com/docs/integrations. The implementation is mnki.adapters.langchain in the mnki package; this distribution is its LangChain-facing name. Apache-2.0.