mcp-highlighter

What it is

A distributed system consisting of a Chrome extension, an HTTP bridge client, and an MCP storage server. It allows users to highlight or identify text formatted with specific [MCP-START]/[MCP-END] tags on any webpage, capture that content, and store it in a persistent JSON file or S3 bucket for retrieval by Model Context Protocol-compatible AI assistants like Claude.

Features

• Strict parsing of [MCP-START]/[MCP-END] delimited text blocks from web pages
• Chrome extension with visual status indicators and manual scan triggers
• HTTP bridge client running on localhost:3001 to receive extension payloads
• Persistent storage via local JSON files or AWS S3 buckets
• Full MCP compliance allowing AI models to list and read stored memories
• Deduplication logic using content hashing to prevent duplicate entries

Quickstart

cd mcp-servers/memory-server && npm install && npm run build
cd ../../mcp_client && npm install && npm run build
npm start
cd ../extension
# Load unpacked extension in chrome://extensions

Architecture

flowchart TD
    User((User)) -->|Highlights Text| Browser[Chrome Browser]
    Browser -->|DOM Scan| Ext[Chrome Extension]
    Ext -->|HTTP POST /mcp/blocks| Bridge[MCP Client Bridge :3001]
    Bridge -->|JSON-RPC / Stdio| MCPServer[MCP Storage Server]
    MCPServer -->|Write| LocalStore[(Local JSON File)]
    MCPServer -->|Write| S3Store[(AWS S3 Bucket)]
    AI[Claude / LLM] -->|MCP Protocol| MCPServer

How it's built

The system is built with TypeScript across three distinct components: a Chrome Extension (content scripts) for DOM scanning, an Express.js HTTP server acting as a bridge/client, and an MCP SDK-based server for standardized tool/resource exposure. It uses strict regex parsing for block detection and JSON-RPC/Stdio for MCP communication.

How it runs

sequenceDiagram
    participant U as User
    participant E as Chrome Extension
    participant B as MCP Client Bridge
    participant S as MCP Storage Server
    participant D as Data Store

    U->>E: Visits page with [MCP-START] block
    E->>E: scanMCP() regex parse
    E->>B: POST /mcp/blocks (JSON payload)
    B->>B: Validate & Hash content
    B->>S: Forward memory entry
    S->>D: Append to memories.json
    S-->>B: Acknowledge storage
    B-->>E: 200 OK
    E->>U: Update UI status 'Found 1'

How to apply & reuse

Use this project to give stateful memory capabilities to stateless LLM sessions. Developers can extend the storage backend (S3/File), modify the parsing logic for different markup formats, or integrate the MCP server into custom AI agent workflows requiring persistent user context.

At a glance