AI Integration

feedpipe can pipe reviewer feedback directly to AI coding agents, enabling automated analysis and fixes.

Supported Agents

feedpipe supports several AI coding agents out of the box:

Agent	Command	Description
`claude`	`claude`	Anthropic’s Claude CLI
`codex`	`codex`	OpenAI Codex CLI
`gemini`	`gemini`	Google Gemini CLI
`opencode`	`opencode`	OpenCode CLI
`custom`	(your command)	Any custom agent

Basic Usage

feedpipe 3000 --ai claude

When a reviewer adds a comment, it’s automatically sent to Claude:

  💬 Comment from Reviewer
     /checkout
     Element: .price-total
     "This should show the discounted price"

  AI agent: claude
  AI agent connected

Event Format

Events are sent to the AI agent’s stdin as JSON, one per line:

{
  "type": "feedback",
  "source": "feedpipe",
  "tunnel": {
    "url": "https://abc123.feedpipe.dev",
    "slug": "abc123"
  },
  "annotation": {
    "id": "ann_123",
    "type": "pin_comment",
    "comment": "This should show the discounted price",
    "pagePath": "/checkout",
    "element": {
      "selector": ".price-total",
      "innerText": "$99.00"
    }
  },
  "reviewer": "Reviewer 42",
  "timestamp": "2026-01-25T10:30:15.000Z"
}

Custom Agents

Use --ai custom with --ai-command for custom agents:

feedpipe 3000 --ai custom --ai-command "node my-agent.js"

Your agent receives events on stdin and can process them however you like.

Example Custom Agent

import readline from 'readline';

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

rl.on('line', (line) => {
  const event = JSON.parse(line);

  console.log(`Received feedback: ${event.annotation.comment}`);
  console.log(`On element: ${event.annotation.element?.selector}`);
  console.log(`Page: ${event.annotation.pagePath}`);

  // Process the feedback...
});

Filtering Events

By default, all annotation events are piped. Use --ai-filter to limit:

# Only pipe comments, not highlights
feedpipe 3000 --ai claude --ai-filter pin_comment

# Only pipe highlights
feedpipe 3000 --ai claude --ai-filter rect_highlight

# Both (default)
feedpipe 3000 --ai claude --ai-filter pin_comment,rect_highlight

Working Directory

Set the working directory for the AI agent:

feedpipe 3000 --ai claude --ai-cwd /path/to/project

This is useful when the AI agent needs to access project files.

Best Practices

1. Use Specific Selectors

The more specific the element selector, the better the AI can understand context:

{
  "element": {
    "selector": "form.checkout button[type='submit']",
    "innerText": "Place Order"
  }
}

2. Write Clear Comments

Reviewers should write actionable comments:

❌ “This is wrong”
✅ “The price should show the 20% discount applied”

3. Include Context

The AI receives the page path and element, but additional context in comments helps:

❌ “Fix this”
✅ “On mobile, this button is too small to tap easily”

4. Filter Noise

If you’re getting too many events, use --ai-filter to focus on what matters.