How I Built an AI Agent That Earns $150 Reviewing Pull Requests

ZG361 — Thu, 09 Apr 2026 12:21:27 +0000

How I Built an AI Agent That Earns $150 Reviewing Pull Requests

What if your code could write code that earns money? That's exactly what I built: a PR Review Agent that autonomously reviews GitHub pull requests and claims bounties in cryptocurrency.

In this article, I'll walk you through how I built claude-review-agent, a CLI tool that fetches GitHub PR diffs, generates structured code reviews, and earned me a $150 bounty.

The Problem

Open source maintainers are overwhelmed. PRs pile up, reviews take hours, and contributors wait days for feedback. What if an AI could handle the first-pass review?

RustChain, a DePIN blockchain project, had a bounty for exactly this:

"Build a CLI tool that reviews PRs and posts structured comments"

Reward: $150 in RTC tokens

The Solution: claude-review-agent

I built a Node.js CLI tool that:

Fetches PR diffs via GitHub API
Generates structured reviews using Claude AI
Posts comments directly on PRs
Outputs Markdown for easy reading

Architecture

┌─────────────┐     ┌──────────────┐     ┌─────────────┐
│ GitHub API  │────▶│   claude-    │────▶│   Claude    │
│  (PR diff)  │     │  review.js   │     │     AI      │
└─────────────┘     └──────────────┘     └─────────────┘
                           │
                           ▼
                    ┌──────────────┐
                    │   Markdown   │
                    │   Review     │
                    └──────────────┘

Step 1: Parse the PR URL

First, we need to extract owner, repo, and PR number from the URL:

function parsePrUrl(url) {
  const match = url.match(/github\.com\/([^/]+)\/([^/]+)\/pull\/(\d+)/);
  if (!match) throw new Error('Invalid GitHub PR URL');
  return {
    owner: match[1],
    repo: match[2],
    prNumber: parseInt(match[3])
  };
}

Step 2: Fetch the PR Diff

Using the GitHub API with Accept: application/vnd.github.v3.diff header:

function fetchPrDiff(owner, repo, prNumber, token) {
  const diff = execSync(
    `curl -s -L -H "Authorization: Bearer ${token}" ` +
    `-H "Accept: application/vnd.github.v3.diff" ` +
    `"https://api.github.com/repos/${owner}/${repo}/pulls/${prNumber}"`,
    { encoding: 'utf-8', maxBuffer: 10 * 1024 * 1024 }
  );
  return diff;
}

Step 3: Generate the Review

Here's where Claude AI shines. We send the diff and ask for structured analysis:

const prompt = `You are an expert code reviewer. Analyze this PR diff and provide:

## Summary
Brief overview of changes (2-3 sentences)

## Risks
List potential issues or concerns (bulleted list)

## Suggestions
Improvement recommendations (bulleted list)

## Confidence
Rate your confidence: High/Medium/Low

Diff:
${diff}`;

const response = await anthropic.messages.create({
  model: 'claude-3-5-sonnet-20241022',
  max_tokens: 4096,
  messages: [{ role: 'user', content: prompt }]
});

Step 4: Output Formats

The tool supports multiple output modes:

# Print to console
claude-review --pr https://github.com/owner/repo/pull/123

# Save to file
claude-review --pr https://github.com/owner/repo/pull/123 --output review.md

# Post as PR comment
claude-review --pr https://github.com/owner/repo/pull/123 --post

Example Output

Here's a real review from testing on awesome-depin PR #28:

## Summary
This PR adds RustChain to the DePIN infrastructure list. It's a minimal 
addition of one entry under the L1/L2 category with project description 
and links.

## Risks
- Low complexity change, minimal risk
- Link validation should be verified

## Suggestions
- Consider adding a brief note about the Proof-of-Antiquity mechanism
- The description could mention supported hardware architectures

## Confidence
Medium

Testing on Real PRs

I tested the agent on two real pull requests:

PR	Repository	Result
#28	iotexproject/awesome-depin	✅ Review generated
#2372	rust-unofficial/awesome-rust	✅ Review generated

Both tests produced meaningful, actionable feedback.

The Bounty Submission

After testing, I submitted the tool to the RustChain bounty issue:

Repository: https://github.com/ZG361/claude-review-agent

Acceptance criteria met:

✅ Works via CLI

✅ Structured Markdown output

✅ Tested on 2 real PRs

✅ README with setup/usage instructions

Result: $150 bounty claimed!

Installation

Want to use it yourself?

# Clone the repo
git clone https://github.com/ZG361/claude-review-agent
cd claude-review-agent

# Install dependencies
npm install

# Set environment variables
export GH_TOKEN=your_github_token
export ANTHROPIC_API_KEY=your_anthropic_key

# Run
node bin/claude-review.js --pr https://github.com/owner/repo/pull/123

Lessons Learned

Start with the acceptance criteria - Read the bounty requirements carefully before coding
Test on real data - Use actual PRs to validate your tool works
Template mode is useful - When you don't have API access, generate structured templates
Document everything - A good README is half the battle

What's Next?

The agent is just the beginning. Future improvements could include:

Multi-language support for non-English reviews
Custom review templates per repository
Integration with CI/CD for automated reviews on every PR
Support for other AI providers (GPT-4, Gemini, etc.)

Conclusion

Building AI agents that earn cryptocurrency isn't science fiction—it's happening right now. The claude-review-agent took about 3 hours to build and earned $150 in RTC tokens.

The future of work is AI-assisted, and platforms like RustChain are creating the infrastructure for AI agents to participate in the economy.

Your old hardware might be worth more than you think—and so is your code.

Built with ❤️ and Claude AI. Check out RustChain for more bounties.

Wallet: ZG361

DEV Community: ZG361

How I Built an AI Agent That Earns $150 Reviewing Pull Requests

How I Built an AI Agent That Earns $150 Reviewing Pull Requests

The Problem

The Solution: claude-review-agent

Architecture

Step 1: Parse the PR URL

Step 2: Fetch the PR Diff

Step 3: Generate the Review

Step 4: Output Formats

Example Output

Testing on Real PRs

The Bounty Submission

Installation

Lessons Learned

What's Next?

Conclusion