Framework integrations (LangChain, LangGraph, CrewAI, AutoGen)

AgentNexusAPI is HTTP-first (POST /api/v1/evaluate, POST /api/v1/receipt). For MCP-native hosts (Cursor, Claude Desktop, custom agents), use POST /api/mcp JSON-RPC — see MCP_GATEWAY.md. Example configs and a stdio bridge are in integrations/mcp-stdio-bridge/. Official thin clients live in-repo so you can wrap any framework in a few lines.

Client	Location	Install (until published)
Python	integrations/python/agentnexus	pip install ./integrations/python/agentnexus from repo root
TypeScript	integrations/typescript/agentnexus-sdk	npm run build in that folder, then npm link or path dep

Set AGENTNEXUS_API_KEY and AGENTNEXUS_BASE_URL (your deployed origin, no trailing slash) in the environment.

“Three lines” core pattern

After install, governance is: construct client → call evaluate → branch on status (then receipt when execution completes).

Python

from agentnexus import AgentNexus
gx = AgentNexus()  # uses AGENTNEXUS_API_KEY + AGENTNEXUS_BASE_URL
out = gx.evaluate("my_policy", {"payload_scope": {"intent": "send_email"}})

TypeScript

import { AgentNexus } from "@agentnexus/sdk";
const gx = new AgentNexus();
const out = await gx.evaluate("my_policy", {
	payload_scope: { intent: "send_email" },
});

Framework-specific snippets below add one indirection (Runnable, graph node, crew step, agent hook) around the same calls.

---

LangChain (Python)

Use a RunnableLambda (or wrap evaluate inside any custom tool). Requires langchain-core.

from langchain_core.runnables import RunnableLambda
from agentnexus import AgentNexus

gx = AgentNexus()

def check_governance(input: dict) -> dict:
    return gx.evaluate("my_policy", {"payload_scope": input})

chain = RunnableLambda(check_governance) | your_downstream_chain

---

LangGraph (Python)

Call the client inside a graph node before side-effect nodes.

from agentnexus import AgentNexus

gx = AgentNexus()

def governance_node(state: dict) -> dict:
    r = gx.evaluate("my_policy", {"payload_scope": state.get("payload", {})})
    return {**state, "governance": r}

---

CrewAI (Python)

Invoke evaluate at the start of a task or in a before_kickoff callback (exact API depends on your CrewAI version).

from agentnexus import AgentNexus

gx = AgentNexus()

def before_agent_act(context: dict) -> None:
    r = gx.evaluate("my_policy", {"payload_scope": context})
    if r.get("status") not in ("auto_approved", "approved"):
        raise RuntimeError("blocked by policy")

---

AutoGen (Python)

Call evaluate from a user-proxy hook or before registering a tool that performs external actions.

from agentnexus import AgentNexus

gx = AgentNexus()

def pre_tool_call(tool_input: dict) -> None:
    gx.evaluate("my_policy", {"payload_scope": tool_input})

---

LangChain.js / LangGraph.js (TypeScript)

Use @agentnexus/sdk with RunnableLambda from @langchain/core/runnables (same idea as Python).

import { RunnableLambda } from "@langchain/core/runnables";
import { AgentNexus } from "@agentnexus/sdk";

const gx = new AgentNexus();
const governed = RunnableLambda.from(async (input: Record<string, unknown>) =>
	gx.evaluate("my_policy", { payload_scope: input }),
);

---

Publishing

• PyPI: package name agentnexus may need conflict check; bump version in pyproject.toml and publish from integrations/python/agentnexus.
• npm: @agentnexus/sdk — run npm run build, then npm publish from integrations/typescript/agentnexus-sdk.

Until then, install from path / git as in the table above.

---

Roadmap

• Optional peer packages (langchain-agentnexus, etc.) if the community wants zero-boilerplate imports.
• CI: smoke tests for clients against a mocked HTTP server.

OpenAPI reference: public/openapi.yaml.