FACT·O·CHECKER 🔍

Autonomous OSINT & Narrative Intelligence Agent 🤖

Fact·O·Checker is a production-grade news verification system that combines deterministic infrastructure analysis with an autonomous AI reasoning agent. Paste any news headline, claim, or URL and receive a structured verdict backed by real-time web search, domain reputation data, and WHOIS registration analysis.

✨ How It Works: Hybrid Agentic Architecture

Fact·O·Checker uses a deliberate two-phase architecture that maximises reliability without sacrificing autonomy.

⚙️ Phase 1 — Deterministic Pre-Processing

When the input contains a URL, the system handler immediately extracts the domain and fires three infrastructure checks in parallel, before the AI agent is even initialised:

URL input detected
       │
       ▼
extractDomain()  ──────────────────────────────────────────────────
       │                                                           │
       ├── fetchVirusTotal(domain)  ← VirusTotal API v3            │
       ├── fetchWhois(domain)       ← RDAP (ICANN standard)        │  Promise.all
       └── ──────────────────────────────────────────────────────► │
                                                                   │
                    Results injected into agent prompt ◄───────────┘

These checks are always executed when a URL is present — they do not depend on the AI model deciding to call them. This eliminates a key failure mode in purely agentic systems: model discretion over whether to invoke tools.

The raw JSON results from VirusTotal and WHOIS are injected verbatim into the user prompt so the agent works with exact, unmodified data:

--- Infrastructure Analysis (use these exact values for flags) ---
WHOIS data: {"domain":"elpais.com","registrar":"Ascio Technologies","ageDays":9221,...}
VirusTotal data: {"malicious":0,"harmless":62,"categories":["news and media"],...}
---

⚙️ Phase 2 — Agentic Synthesis (gpt-oss-120b via Groq)

With the infrastructure data already in context, the agent focuses on what it does best: autonomous web research and narrative analysis.

The agent runs a ReAct (Reasoning + Acting) loop using the Vercel AI SDK's generateText with maxSteps: 5. In each step it can:

Formulate and execute Tavily search queries
Reformulate queries if results are insufficient
Cross-reference multiple sources
Detect fact-checker coverage (Maldita, Snopes, AFP, etc.)
Synthesise all evidence into a structured JSON verdict

The agent decides autonomously how many searches to run, what angles to investigate, and when it has enough evidence to deliver a verdict.

🔀 Why This Hybrid Approach

Concern	Pure Agentic	Pure Deterministic	Hybrid (Fact·O·Checker)
VT + WHOIS always run	✗ Model may skip	✓	✓
Web research quality	✓ Autonomous	✗ Fixed queries	✓ Autonomous
Token efficiency	✗ Wastes steps on infra	—	✓ Infra pre-loaded
Hallucinated flags	✗ Risk	✓ None	✓ Data injected
Adaptability	✓	✗	✓

The model is still the reasoning engine. It decides what to search, how many times, and what conclusion to draw. It just receives infrastructure data as a given rather than having to fetch it itself.

⚖️ Verdict System

Each verification returns one of four verdicts:

Verdict	Meaning	Color
REAL	Multiple credible sources corroborate the claim	Sky blue
FAKE	Claim contradicted by credible sources or satire detected	Red
SUSPICIOUS	Domain < 30 days old, conflicting sources, or no credible corroboration	Amber
UNVERIFIABLE	No sources found; claim cannot be confirmed or denied	Gray

Alongside the verdict: a confidence score (0–100), a plain-language summary in the input language, source links, and structured flags.

🚩 Flag System

Flags are colour-coded by severity:

** 🟢 Green — informational/positive**

Domain registered X days ago (since YEAR)
Registrar: [name]
Registrant country: [code]
VirusTotal: clean — 0 malicious, X harmless votes
Hosting category: [news / media / etc]
Fact-checked by: maldita.es (detected from Tavily results)

** 🟡 Amber — caution**

Conflicting source quality
Unverifiable claims

** 🔴 Red — critical**

Domain registered less than 30 days ago ⚠️
VirusTotal: X malicious votes detected ⚠️

🛠️ Stack

Layer	Technology	Notes
Framework	Next.js 14 (App Router)	Serverless-first
AI Reasoning	gpt-oss-120b	Via Groq — fast inference, generous free tier
AI SDK	Vercel AI SDK (`ai@4.3.19`)	`generateText` + `maxSteps` ReAct loop
Web Search	Tavily Search API	Optimised for LLM agents; 1,000 searches/month free
Domain Reputation	VirusTotal API v3	500 req/day free
WHOIS / Registration	RDAP (ICANN)	HTTP standard, no API key required
Bot Protection	hCaptcha (invisible)	Serverside token verification
UI	shadcn/ui + Tailwind CSS v4	Modernist Glassmorphic Light design
Font	Inter (Google Fonts)	Weights 300 / 500 / 700 / 900
Deployment	Vercel	60s function timeout configured

</> Getting Started

Prerequisites

Node.js 18+
Accounts for: Groq, Tavily, VirusTotal, hCaptcha

📦 Installation

git clone https://github.com/aisurf3r/Fact-o-checker.git
cd Fact-o-checker
npm install

🗝️ Environment Variables

Create .env.local in the project root:

GROQ_API_KEY=           # Groq console → API Keys
TAVILY_API_KEY=         # Tavily dashboard → API
VIRUSTOTAL_API_KEY=     # VirusTotal → API Key
HCAPTCHA_SECRET_KEY=    # hCaptcha dashboard → Secret Key

🚀 Run Locally

npm run dev

Open http://localhost:3000.

Note: 📝 The agent makes outbound HTTP requests to Groq, Tavily, VirusTotal and RDAP. These calls require a real server environment. The Bolt/StackBlitz WebContainer may block external API calls due to CORS related behaviour — deploy to server enviroment for full functionality.

API Reference

`POST /api/verify`

Request body:

{
  "input": "https://example.com/article  OR  plain text claim",
  "hcaptchaToken": "token-from-frontend-widget"
}

Response:

{
  "verdict": {
    "verdict": "REAL | FAKE | SUSPICIOUS | UNVERIFIABLE",
    "confidence": 85,
    "summary": "Plain-language explanation in the input language.",
    "sources": ["https://...", "https://..."],
    "flags": ["Domain registered 9221 days ago (since 2001)", "VirusTotal: clean — 0 malicious, 62 harmless votes"]
  },
  "steps": 3,
  "infraChecked": true,
  "domain": "elpais.com"
}

Fields:

infraChecked — true if input was a URL and VT + WHOIS were executed
domain — extracted hostname (present when infraChecked is true)
steps — number of ReAct reasoning steps the agent completed

🆓 Free Tier Limits

Service	Free Tier	Bottleneck
Groq (gpt-oss-120b)	1,000 req/day	Not critical for personal use
Tavily	1,000 searches/month	Primary bottleneck (~2 searches/verification = ~500 verifications/month)
VirusTotal	500 req/day	Not critical
RDAP	Unlimited	No key required
hCaptcha	1M verifications/month	Not critical

🏗️ Project Structure

app/
  api/verify/
    route.ts          POST endpoint — pre-processing + agent
  layout.tsx          Root layout, metadata, fonts
  page.tsx            UI: Idle / Loading / Result states
  globals.css         Tailwind base + custom animations
  favicon.ico         Site icon

🎨 Design System

The interface follows a Modernist Glassmorphic Light aesthetic:

Background:   linear-gradient(135deg, #e0f2fe, #f0fdf4, #f8fafc)
Cards:        rgba(255,255,255,0.72) + backdrop-filter: blur(20px)
Accent A:     #0EA5E9  (sky blue  — primary actions, REAL verdict)
Accent B:     #10B981  (emerald   — success states, positive flags)
Danger:       #EF4444  (red       — FAKE verdict, critical flags)
Warning:      #F59E0B  (amber     — SUSPICIOUS, caution flags)
Typography:   Inter 300 / 500 / 700 / 900
Monospace:    Consolas (data fields, domain names, API values)

🌐 Roadmap

Phase	Status	Description
V1 — Core	✅ Live	Hybrid agent: deterministic infra + agentic web search
V2 — Cache	Planned	Supabase cache layer — `hash(input)` → cached verdict, zero tokens on repeated queries
V3 — Global Watch	Planned	Mapbox dashboard — domain IP geolocation, disinformation heatmaps, `flyTo()` animations
Scale	Future	Enterprise API, rate limiting, usage analytics

📝 License

MIT

_{Developer: aisurf3r · Brain: gpt-oss-120b via Groq · Contact: aisurf3r@gmail.com}

Name		Name	Last commit message	Last commit date
app		app
README.md		README.md
components.json		components.json
next-env.d.ts		next-env.d.ts
next.config.mjs		next.config.mjs
package.json		package.json
postcss.config.mjs		postcss.config.mjs
tsconfig.json		tsconfig.json

Navigation Menu

Repository files navigation

FACT·O·CHECKER 🔍

✨ How It Works: Hybrid Agentic Architecture

⚙️ Phase 1 — Deterministic Pre-Processing

⚙️ Phase 2 — Agentic Synthesis (gpt-oss-120b via Groq)

🔀 Why This Hybrid Approach

⚖️ Verdict System

🚩 Flag System

🛠️ Stack

</> Getting Started

Prerequisites

📦 Installation

🗝️ Environment Variables

🚀 Run Locally

API Reference

`POST /api/verify`

🆓 Free Tier Limits

🏗️ Project Structure

🎨 Design System

🌐 Roadmap

📝 License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

FACT·O·CHECKER 🔍

✨ How It Works: Hybrid Agentic Architecture

⚙️ Phase 1 — Deterministic Pre-Processing

⚙️ Phase 2 — Agentic Synthesis (gpt-oss-120b via Groq)

🔀 Why This Hybrid Approach

⚖️ Verdict System

🚩 Flag System

🛠️ Stack

</> Getting Started

Prerequisites

📦 Installation

🗝️ Environment Variables

🚀 Run Locally

API Reference

POST /api/verify

🆓 Free Tier Limits

🏗️ Project Structure

🎨 Design System

🌐 Roadmap

📝 License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

`POST /api/verify`