Run a LangGraph node in a sandbox
Keep orchestration in LangGraph and run commands in an isolated machine. This graph computes a total inside a sandbox, then checks it in a second node. It uses no model calls.
Run the graph
Set TANGLE_API_KEY from your API keys.
Save the program below as sandbox-graph.mts and install its dependencies:
npm install @tangle-network/sandbox @langchain/langgraph @langchain/core
npm install --save-dev tsx
npx tsx sandbox-graph.mtsimport assert from "node:assert/strict";
import { Annotation, END, START, StateGraph } from "@langchain/langgraph";
import { Sandbox } from "@tangle-network/sandbox";
const client = new Sandbox({
apiKey: process.env.TANGLE_API_KEY!,
baseUrl: "https://sandbox.tangle.tools",
});
const State = Annotation.Root({
total: Annotation<number>(),
verified: Annotation<boolean>(),
});
const box = await client.create({ agent: false, maxLifetimeSeconds: 180 });
try {
const graph = new StateGraph(State)
.addNode("calculate", async () => {
const result = await box.exec("python3 -c 'print(sum([120, 80, 45, 15]))'");
assert.equal(result.exitCode, 0, result.stderr);
const total = Number(result.stdout.trim());
assert(Number.isFinite(total));
return { total };
})
.addNode("verify", (state) => {
assert.equal(state.total, 260);
return { verified: true };
})
.addEdge(START, "calculate")
.addEdge("calculate", "verify")
.addEdge("verify", END)
.compile();
const result = await graph.invoke({ total: 0, verified: false });
assert.equal(result.verified, true);
console.log(JSON.stringify(result));
} finally {
await box.delete();
}The result is {"total":260,"verified":true}. Replace the calculation with your
own command, or call box.prompt() from a node to run a coding agent.
LangGraph owns the graph state and edges; see its Annotation reference. For a graph where each worker is a coding agent, see Agent graphs.