§ integrations / langgraph

LangGraph · node-level scope enforcement, graph replay.

LangGraph's stateful graphs deserve the same audit posture as their stateless cousins. Fleet's adapter scope-checks every tool node and emits agent-action envelopes per execution — plus a graph-replay tool that re-derives the full audit trail from the persisted graph state.

· shipping· published to npm · MIT· package @orangecheck/agent-langgraph@0.2.0

What it does

LangGraph composes long-running stateful agents from typed nodes. The adapter touches three places:

Tool nodes — scope check before the node runs; envelope emission after. Same agent-action envelope shape as every other adapter.
Graph state checkpoints — checkpoint hashes are folded into the audit bundle so a verifier can prove which graph state the agent was operating from when an action fired.
Replay tool — given a persisted graph state and the audit bundle, re-derive every action receipt and confirm the bundle's claimed history matches what the graph would have produced.

Why this is harder than a stateless adapter

LangGraph agents persist state across many turns. The compliance question becomes: at what graph state did this agent take this action? The adapter has to bind action envelopes not just to the input arguments but to the graph state hash. This is a real engineering surface, which is why it sits below MCP / Anthropic / OpenAI / Vercel AI in priority order.

Status

Adapter package: @orangecheck/agent-langgraph@0.2.0 is on npm. Source in oc-packages/agent-langgraph. MIT. Ships stampToolCall, postActionToFleet, and the high-level ocToolNode({verb, execute}) wrapper for tool-node execution. Graph state hashing + the dedicated replay verifier remain on the roadmap behind a design partner ask — the v0.2.0 surface focuses on the per-tool-call audit envelope.

What Fleet adds on top: the managed signing pipeline, OC Stamp anchoring, Nostr publishing, audit-bundle assembly, and the operator dashboard.

preview · concept sketch

import { StateGraph } from '@langchain/langgraph';
import { ocToolNode } from '@orangecheck/agent-langgraph';

const createInvoice = ocToolNode({
  verb:    'invoice.create',
  execute: async (args) => createImpl(args),
});

const graph = new StateGraph(MyState)
  .addNode('createInvoice', async (state) => {
    const { result, posted } = await createInvoice.execute(state.invoice, {
      agent, delegation, callId: state.runId,
      fleet: {
        apiToken: process.env.OC_FLEET_TOKEN!,
        projectId: process.env.OC_FLEET_PROJECT_ID!,
      },
    });
    return { ...state, lastReceipt: posted?.id };
  })
  // … sendNotification, conditional branches, etc.

// each ocToolNode call emits an OC Agent action envelope, posted to
// fleet for OC Stamp anchoring + the audit bundle. graph-state hashing
// + replay verification land in a future minor when a design partner
// pulls them forward.

Already using the package?

Fleet adds the managed pipeline (OC Stamp anchoring, Nostr publishing, audit-bundle assembly, operator dashboard) on top of the open package. Design partners pair with an OrangeCheck engineer for the integration.

request access →read the docs →back to gallery →