Guide: Human-in-the-loop approval with Anthropic

Open in

This guide shows you how to implement a human-in-the-loop (HITL) approval workflow for AI agent tool calls using Anthropic and Ably. The agent requests human approval before executing sensitive operations, with role-based access control to verify approvers have sufficient permissions.

When the model calls a tool that requires human approval, the tool implementation itself handles the approval check before executing. Rather than executing immediately, the tool publishes an approval-request message to an Ably channel, waits for an approval-response from a human approver, verifies the approver has the required role using claims embedded in their JWT token, and only then executes the action. The model calls the tool as normal, and the approval logic lives inside the tool's implementation.

Prerequisites

To follow this guide, you need:

Node.js 20 or higher
An Anthropic API key
An Ably API key

Useful links:

Create a new NPM package, which will contain the agent, client, and server code:

mkdir ably-anthropic-hitl-example && cd ably-anthropic-hitl-example
npm init -y

Install the required packages using NPM:

npm install @anthropic-ai/sdk@^0.71 ably@^2 express jsonwebtoken

Export your API keys to the environment:

export ANTHROPIC_API_KEY="your_anthropic_api_key_here"
export ABLY_API_KEY="your_ably_api_key_here"

Step 1: Initialize the agent

Set up the agent that will call Anthropic and request human approval for sensitive operations. This example uses a publish_blog_post tool that requires authorization before execution.

Initialize the Anthropic and Ably clients, and create a channel for communication between the agent and human approvers.

Add the following to a new file called agent.mjs:

JavaScript

import Anthropic from '@anthropic-ai/sdk';
import Ably from 'ably';

const anthropic = new Anthropic();

// Initialize Ably Realtime client
const realtime = new Ably.Realtime({
  key: process.env.ABLY_API_KEY,
  echoMessages: false
});

// Wait for connection to be established
await realtime.connection.once('connected');

// Create a channel for HITL communication
const channel = realtime.channels.get('ai:sit-gap-cog');

// Track pending approval requests
const pendingApprovals = new Map();

// Function that executes the approved action
async function publishBlogPost(args) {
  const { title } = args;
  console.log(`Publishing blog post: ${title}`);
  // In production, this would call your CMS API
  return { published: true, title };
}

Tools that modify data, access sensitive resources, or perform actions with business impact are good candidates for HITL approval workflows.

Step 2: Request human approval

When the model returns a tool use block, publish an approval request to the channel and wait for a human decision. The tool use ID is passed in the message headers to correlate requests with responses.

Add the approval request function to agent.mjs:

JavaScript

async function requestHumanApproval(toolUse) {
  const approvalPromise = new Promise((resolve, reject) => {
    pendingApprovals.set(toolUse.id, { toolUse, resolve, reject });
  });

  await channel.publish({
    name: 'approval-request',
    data: {
      tool: toolUse.name,
      arguments: toolUse.input
    },
    extras: {
      headers: {
        toolCallId: toolUse.id
      }
    }
  });

  console.log(`Approval request sent for: ${toolUse.name}`);
  return approvalPromise;
}

The toolUse.id provided by Anthropic correlates the approval request with the response, enabling the agent to handle multiple concurrent approval flows.

Set up a subscription to receive approval decisions from human users. When a response arrives, verify the approver has sufficient permissions using role-based access control before resolving the pending promise.

Add the subscription handler to agent.mjs:

JavaScript

async function subscribeApprovalResponses() {
  // Define role hierarchy from lowest to highest privilege
  const roleHierarchy = ['editor', 'publisher', 'admin'];

  // Define minimum role required for each tool
  const approvalPolicies = {
    publish_blog_post: { minRole: 'publisher' }
  };

  function canApprove(approverRole, requiredRole) {
    const approverLevel = roleHierarchy.indexOf(approverRole);
    const requiredLevel = roleHierarchy.indexOf(requiredRole);
    return approverLevel >= requiredLevel;
  }

  await channel.subscribe('approval-response', async (message) => {
    const { decision } = message.data;
    const toolCallId = message.extras?.headers?.toolCallId;
    const pending = pendingApprovals.get(toolCallId);

    if (!pending) {
      console.log(`No pending approval for tool call: ${toolCallId}`);
      return;
    }

    const policy = approvalPolicies[pending.toolUse.name];
    // Get the trusted role from the JWT user claim
    const approverRole = message.extras?.userClaim;

    // Verify the approver's role meets the minimum required
    if (!canApprove(approverRole, policy.minRole)) {
      console.log(`Insufficient role: ${approverRole} < ${policy.minRole}`);
      pending.reject(new Error(
        `Approver role '${approverRole}' insufficient for required '${policy.minRole}'`
      ));
      pendingApprovals.delete(toolCallId);
      return;
    }

    // Process the decision
    if (decision === 'approved') {
      console.log(`Approved by ${approverRole}`);
      pending.resolve({ approved: true, approverRole });
    } else {
      console.log(`Rejected by ${approverRole}`);
      pending.reject(new Error(`Action rejected by ${approverRole}`));
    }
    pendingApprovals.delete(toolCallId);
  });
}

The message.extras.userClaim contains the role embedded in the approver's JWT token, providing a trusted source for authorization decisions. See user claims for details on embedding claims in tokens. This ensures only users with sufficient privileges can approve sensitive operations.

Step 4: Process tool calls

Create a function to process tool use blocks by requesting approval and executing the action if approved.

Add the tool processing function to agent.mjs:

JavaScript

async function processToolUse(toolUse) {
  if (toolUse.name === 'publish_blog_post') {
    // requestHumanApproval returns a promise that resolves when the human
    // approves the tool use, or rejects if the human explicitly rejects
    // the tool call or the approver's role is insufficient.
    await requestHumanApproval(toolUse);
    return await publishBlogPost(toolUse.input);
  }
  throw new Error(`Unknown tool: ${toolUse.name}`);
}

The function awaits approval before executing. If the approver rejects or has insufficient permissions, the promise rejects and the tool is not executed.

Step 5: Run the agent

Create the main agent loop that sends prompts to the model and processes any tool use blocks that require approval.

Add the agent runner to agent.mjs:

JavaScript

async function runAgent(prompt) {
  await subscribeApprovalResponses();

  console.log(`User: ${prompt}`);

  const response = await anthropic.messages.create({
    model: 'claude-sonnet-4-5',
    max_tokens: 1024,
    tools: [
      {
        name: 'publish_blog_post',
        description: 'Publish a blog post to the website. Requires human approval.',
        input_schema: {
          type: 'object',
          properties: {
            title: {
              type: 'string',
              description: 'Title of the blog post to publish'
            }
          },
          required: ['title']
        }
      }
    ],
    messages: [{ role: 'user', content: prompt }]
  });

  const toolUseBlocks = response.content.filter(block => block.type === 'tool_use');

  for (const toolUse of toolUseBlocks) {
    console.log(`Tool use: ${toolUse.name}`);
    try {
      const result = await processToolUse(toolUse);
      console.log('Result:', result);
    } catch (err) {
      console.error('Tool use failed:', err.message);
    }
  }
}

runAgent("Publish the blog post called 'Introducing our new API'");

Step 6: Create the authentication server

The authentication server issues JWT tokens with embedded role claims. The role claim is trusted by Ably and included in messages, enabling secure role-based authorization.

Add the following to a new file called server.mjs:

JavaScript

import express from 'express';
import jwt from 'jsonwebtoken';

const app = express();

// Mock authentication - replace with your actual auth logic
function authenticateUser(req, res, next) {
  // In production, verify the user's session/credentials
  req.user = { id: 'user123', role: 'publisher' };
  next();
}

// Return claims to embed in the JWT
function getJWTClaims(user) {
  return {
    'ably.channel.*': user.role
  };
}

app.get('/api/auth/token', authenticateUser, (req, res) => {
  const [keyName, keySecret] = process.env.ABLY_API_KEY.split(':');

  const token = jwt.sign(getJWTClaims(req.user), keySecret, {
    algorithm: 'HS256',
    keyid: keyName,
    expiresIn: '1h'
  });

  res.type('application/jwt').send(token);
});

app.listen(3001, () => {
  console.log('Auth server running on http://localhost:3001');
});

The ably.channel.* claim embeds the user's role in the JWT. When the user publishes messages, this claim is available as message.extras.userClaim, providing a trusted source for authorization.

Run the server:

node server.mjs

Step 7: Create the approval client

The approval client receives approval requests and allows humans to approve or reject them. It authenticates via the server to obtain a JWT with the user's role.

Add the following to a new file called client.mjs:

JavaScript

import Ably from 'ably';
import readline from 'readline';

const rl = readline.createInterface({
  input: process.stdin,
  output: process.stdout
});

const realtime = new Ably.Realtime({
  authCallback: async (tokenParams, callback) => {
    try {
      const response = await fetch('http://localhost:3001/api/auth/token');
      const token = await response.text();
      callback(null, token);
    } catch (error) {
      callback(error, null);
    }
  }
});

realtime.connection.on('connected', () => {
  console.log('Connected to Ably');
  console.log('Waiting for approval requests...\n');
});

const channel = realtime.channels.get('ai:sit-gap-cog');

await channel.subscribe('approval-request', (message) => {
  const request = message.data;

  console.log('\n========================================');
  console.log('APPROVAL REQUEST');
  console.log('========================================');
  console.log(`Tool: ${request.tool}`);
  console.log(`Arguments: ${JSON.stringify(request.arguments, null, 2)}`);
  console.log('========================================');

  rl.question('Approve this action? (y/n): ', async (answer) => {
    const decision = answer.toLowerCase() === 'y' ? 'approved' : 'rejected';

    await channel.publish({
      name: 'approval-response',
      data: { decision },
      extras: {
        headers: {
          toolCallId: message.extras?.headers?.toolCallId
        }
      }
    });

    console.log(`Decision sent: ${decision}\n`);
  });
});

Run the client in a separate terminal:

node client.mjs

With the server, client, and agent running, the workflow proceeds as follows:

The agent sends a prompt to the model that triggers a tool use
The agent publishes an approval request to the channel
The client displays the request and prompts the user
The user approves or rejects the request
The agent verifies the approver's role meets the minimum requirement
If approved and authorized, the agent executes the tool

Next steps

Learn more about human-in-the-loop patterns and verification strategies
Explore identifying users and agents for secure identity verification
Understand sessions and identity in AI-enabled applications
Learn about tool calls for agent-to-agent communication

Guide: Human-in-the-loop approval with Anthropic

Prerequisites

Step 1: Initialize the agent

Step 2: Request human approval

Step 3: Subscribe to approval responses

Step 4: Process tool calls

Step 5: Run the agent

Step 6: Create the authentication server

Step 7: Create the approval client

Next steps