Give a LangChain agent vault access

This example lets a LangChain agent decide when to list and fetch vault secrets. You define custom tools that wrap @1claw/sdk, and the LLM calls them just in time so raw values never sit in the prompt.

1. 1

   Clone the examples repository and install dependencies.

   bash Copy

   git clone https://github.com/1clawAI/1claw-examples.git
   cd 1claw-examples/langchain-agent
   npm install
   cp .env.example .env

2. 2

   Fill in .env with your 1Claw and LLM credentials. npm start uses ONECLAW_API_KEY (user key), while npm run mcp uses ONECLAW_AGENT_API_KEY (agent key).

   text Copy

   ONECLAW_API_KEY=1ck_your_user_key_here
   ONECLAW_VAULT_ID=your-vault-uuid
   ONECLAW_AGENT_API_KEY=ocv_your_agent_key_here
   GOOGLE_API_KEY=your-gemini-key

3. 3

   If your vault is empty, add a test secret with the CLI.

   bash Copy

   1claw secret set demo/hello-world --vault $ONECLAW_VAULT_ID --value "Hello from 1Claw!" --type generic

4. 4

   Note the tool pattern: a DynamicStructuredTool wraps the SDK client so the agent can list secrets. BASE_URL, API_KEY, and VAULT_ID are defined from environment variables at the top of the file.

   typescript Copy

   import { DynamicStructuredTool } from "@langchain/core/tools";
   import { createClient } from "@1claw/sdk";
   import { z } from "zod";
   
   const BASE_URL = process.env.ONECLAW_BASE_URL || "https://api.1claw.co";
   const API_KEY = process.env.ONECLAW_API_KEY!;
   const VAULT_ID = process.env.ONECLAW_VAULT_ID!;
   
   const client = createClient({ baseUrl: BASE_URL, apiKey: API_KEY });
   
   const listSecretsTool = new DynamicStructuredTool({
     name: "list_vault_secrets",
     description: "List all secrets in the vault (metadata only, never values).",
     schema: z.object({}),
     func: async () => {
       const res = await client.secrets.list(VAULT_ID);
       return res.data.secrets.map(s => `${s.path} (${s.type})`).join("\n");
     },
   });

5. 5

   Run the tool-calling agent.

   bash Copy

   npm start

Your LangChain agent now lists vault secrets through the SDK tool and reports paths and types without leaking the value.

What this gives you, and what it does not. Wrapping the SDK in tools lets the model decide when to fetch a secret, which is a real capability and a real widening.

• A tool the model can call is a tool an injection can call. Tool-calling moves the decision from your code to the model, so tool scope becomes the security boundary.
• Returning a value puts it in context. A tool that fetches a secret and returns it has published that secret to the conversation, its history and the provider. Prefer tools that act rather than tools that reveal.
• The published package exists. langchain-1claw ships ready-made tools with names like oneclaw_get_secret and oneclaw_list_secrets, so hand-written wrappers are a choice rather than a requirement.