Fathom

An AI-powered web research agent that autonomously searches Google and YouTube, extracts content, and synthesizes comprehensive reports — with full source traceability and version history.

What It Does

Give Fathom a research topic and it will:

1. Search Google and YouTube for relevant sources
2. Scrape web pages and extract YouTube transcripts
3. Use Claude (via AWS Bedrock) to synthesize findings into a structured report
4. Cite every claim with source IDs linked to the full source content
5. Track changes over time so you can refresh research and see what changed

How It Works

Fathom has two components that work together:

Backend (FastAPI) — Runs the AI agent loop using AWS Bedrock (Claude Sonnet). Manages research projects, the job queue, source storage, and Git-based version history.

Chrome Extension — Acts as the scraping layer. It polls the job queue and executes search/scrape tasks by opening real Chrome tabs in the background. This avoids API keys and rate limits entirely.

User → FastAPI → Bedrock Agent → Job Queue → Chrome Extension → Google/YouTube
                                                                        ↓
User ← Report ← FastAPI ← Markdown + Sources ←────────────────────────┘

Features

• Dual-source research — searches both Google and YouTube in a single session
• Source traceability — every claim is cited ([src-1], [src-2]) and linked to full source content
• Seed URLs — optionally anchor research to specific authoritative sources the agent reads first
• Research refresh — re-run research to find updates; supports partial refresh (e.g. "update the pricing section only")
• Version history — each research project is a Git repo; refreshes create commits with agent-written change summaries
• Diff viewer — side-by-side diffs between any two versions of a report
• Export — download reports as Markdown, plain text, or PDF
• Live activity monitor — real-time panel showing searches, scrapes, and AI thinking as they happen

Tech Stack

Layer	Technology
Backend	Python, FastAPI
Frontend	Vue 3 + Vuetify (CDN, no build step)
LLM	AWS Bedrock (Claude Sonnet 4.6)
Web scraping	crawl4ai (Playwright-based)
Extension	Chrome Manifest V3
Database	SQLite
Report storage	Markdown files + Git (via GitPython)
Diff rendering	diff2html

Prerequisites

• Python 3.10+
• Google Chrome with the Fathom extension installed
• AWS credentials with Bedrock access (Claude Sonnet 4.6 in us-east-1)
• crawl4ai installed and Playwright browsers set up

Setup

1. Install Python dependencies

pip install -r requirements.txt
playwright install

2. Install the Chrome extension

• Open Chrome and go to chrome://extensions
• Enable Developer mode
• Click Load unpacked and select the extension/ folder

3. Start the server

start.bat

Or directly:

uvicorn app.main:app --reload --reload-exclude research --reload-exclude logs

4. Open the app

Go to http://localhost:8000 in Chrome (must be Chrome — the extension only runs there).

Usage

1. Click New Research and enter your topic
2. Optionally add seed URLs (sources the agent should read first)
3. Click Start and watch the live activity panel
4. Once complete, view the report, browse sources, or explore version history
5. Use Refresh later to update the research with new findings

Research Folder

The research/ folder is local-only (excluded from Git). Each project gets its own subfolder with:

• report.md — the generated report
• sources/ — scraped content for each source
• A Git repo for version history

To share research on GitHub, move it to research-public/ which is tracked.

AWS Setup

Fathom uses the AWS SDK with no hardcoded credentials. It works with:

• IAM roles (e.g. on EC2)
• AWS CLI configured credentials (aws configure)
• Environment variables (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY)

Bedrock model access for Claude Sonnet 4.6 must be enabled in the us-east-1 region.