Skip to main content

Self-Improvement Overview

This is the core of recapt: connect your AI coding assistant to behavioral intelligence, and it automatically detects UX issues, writes fixes, and verifies they worked.

What is MCP?​

MCP (Model Context Protocol) is a standard that allows AI coding assistants to connect to external data sources and tools. The recapt MCP server gives your AI assistant access to behavioral data from your session recordings, enabling it to:

  • Find UX issues automatically from real user behavior
  • Investigate root causes with session data
  • Propose fixes based on evidence
  • Track whether fixes actually worked

The Self-Improvement Loop​

The core value of the MCP server is the self-improvement workflow, a closed loop that goes from detecting issues to verifying fixes.

Self-improvement isn't just for bugs

Self-improvement detects all types of UX issues, not just code errors. This includes confusing navigation, unclear CTAs, frustrating form fields, poor page layouts, and any friction that causes users to struggle. If users are having a bad experience, recapt will find it.

PhaseWhat happens
DetectAI finds UX issues from behavioral data (rage clicks, dead clicks, form friction, confusing flows, drop-offs)
DiagnoseAI investigates root cause with session data and element-level friction
FixAI proposes changes (code fixes, UX improvements, copy changes) and you implement them
TrackRecapt captures baseline metrics before the fix goes live
EvaluateAfter deployment, AI compares post-fix metrics to baseline
LearnSuccessful patterns are stored as site knowledge for future reference

Quick Start: The Self-Improvement Skill​

The easiest way to use self-improvement is to install the self-improvement skill, which guides your AI through the complete workflow automatically:

npx @recapt/cli skill install self-improvement

Once installed, just ask:

Heal my site

or

Find and fix UX issues

The AI will run the entire workflow (detecting issues, investigating root causes, proposing fixes, and tracking results) without you having to prompt each step manually.

Under the Hood​

For more control, you can run each step yourself. Here's what a simple manual self-improvement session looks like:

You: "What UX issues do we have?"

AI: I found 3 active issues. The highest priority is rage clicks on the
    checkout button (/checkout page). 47 users experienced this in the
    last 7 days with frustration score of 0.72.

You: "Investigate that checkout issue"

AI: I analyzed the affected sessions and found:
    - Console error: "TypeError: Cannot read property 'submit' of null"
    - The error occurs when users click before the form fully loads
    - Suggested fix: Add a loading state to disable the button until ready

    [Shows code suggestion]

You: "I deployed the fix"

AI: Marked as deployed. I'll track the metrics. Check back in 24-48 hours
    for enough data to evaluate the fix.

You: [24 hours later] "Did the fix work?"

AI: Yes! Comparing the last 24 hours to the baseline:
    - Frustration: 0.72 → 0.31 (↓57%)
    - Rage clicks: 47 → 3 (↓94%)
    - Issue status: Resolved

    The fix was successful. I've stored this pattern for future reference.

This manual approach is useful when you want to:

  • Focus on a specific page or issue
  • Skip certain steps (e.g., triage without fixing)
  • Understand exactly what data the AI is using

Supported IDEs​

The MCP server works with any AI coding assistant that supports the MCP protocol:

  • Cursor
  • Claude Code
  • Windsurf
  • VS Code (with MCP extension)
  • Zed
  • Codex CLI
  • GitHub Copilot

Next Steps​