Most people connect Octoparse MCP to their AI assistant and use it to extract a product list or pull some prices into a table.

That's fine. But I wanted to use it differently.

I wanted Octoparse MCP to act as a live structured intelligence agent — something my AI system could call on demand, in real time, every time a user submits a story or a press angle. Not a batch job. Not a scheduled pipeline. A live tool call that returns clean, structured journalist profiles directly into an LLM scoring engine.

That's what I built with E_MediaScience.

The Problem
PR intelligence tools like Cision and Muck Rack cost $10,000–$30,000 per year. They're inaccessible to most founders, startups, and SMEs — the exact people who most need earned media coverage to grow.

The alternative is hours of manual research: scanning publication mastheads, reading journalist bylines, guessing at beats and tone. Even then, the outreach is generic because there's no structured data behind it.

The core problem isn't finding journalists. It's vocabulary asymmetry.

A founder knows their product. They don't know how a journalist at TechCrunch would classify it, what beat editor covers their category, or which publications have recently covered adjacent topics. Traditional search tools enforce a tight validation loop — keep refining your query until you find something — and most users give up before they get there.

What E_MediaScience Does
E_MediaScience is a multi-tenant earned media operating system. A user submits a story, launch, or campaign brief. The system:

Calls Octoparse MCP with intent-based parameters — not a URL, but a topic and geographic target

Octoparse selects the appropriate journalist discovery template from its 600+ library and executes geo-routed extraction

Returns clean, structured journalist profiles: name, outlet, beat, article history, tone markers, contact data

Feeds that payload directly into Claude for AI newsworthiness scoring and journalist matching

Generates personalised outreach referencing each journalist's actual recent work

Tracks replies, open rates, and campaign strike rate

The entire flow takes under 60 seconds from submission to matched journalist list.

Why Octoparse MCP Changes Everything
Before MCP, my options were:

Build and maintain custom scrapers per publication (brittle, expensive, breaks constantly)

Use a static journalist database (stale, expensive, no real-time beat tracking)

Ask an LLM to find journalists (hallucinated profiles, made-up contact details)

Octoparse MCP eliminates all three problems in a single tool call.

text
User submits: "I've launched an AI video clipping tool for live-sellers"

EMS calls Octoparse MCP:
→ Template: journalist-discovery-tech-ecommerce
→ Parameters: { topic: "AI video tools, live commerce, creator economy", regions: ["UK", "US"] }

Octoparse returns:
→ 12 journalist profiles, structured JSON
→ Beat: "Commerce technology, live shopping, creator tools"
→ Recent articles, outlet, contact data — all clean, no parsing

Claude scores:
→ Newsworthiness: 74/100
→ Top match: [Journalist at The Information, beat: AI/Creator Economy]
→ Personalised pitch: references journalist's last 3 articles
No HTML. No CSS selectors. No fragile extraction logic. The structured payload goes straight into the LLM.

The HungQueryResolver — V1.1 Innovation
The most technically novel part of E_MediaScience is the HungQueryResolver — built specifically around what Octoparse MCP can do when queries fail.

The problem: clients describe their PR targets in natural language that doesn't map cleanly to journalist taxonomies. "Find people who write about the neat tech stuff I make" is a real query. Traditional systems force clarification loops until the user gives up.

The HungQueryResolver uses a Three-Strike Escalation architecture:

Turn 1 — Direct Match
Octoparse MCP called with the raw query. High-confidence matches are returned immediately.

Turn 2 — Drift Validation
If confidence falls below threshold, the user is prompted once for clarification. The system measures whether the new query actually adds new information — or just rephrases the same intent.

Turn 3 — Async Escalation
If the user is circling the same concept in different words, the system stops asking. A background worker fires a broadened Octoparse MCP call with expanded terminology, adjacent industry classifications, and alternate journalist taxonomies — silently, while the UI holds.

Instead of a dead end, the user gets a scored set of alternative matches with a transparent quality rating explaining why each result was surfaced.

This turns a search failure into a consulting asset.

The Multi-Tool MCP Stack
E_MediaScience was built in Cursor IDE using Claude Sonnet and Opus. The full MCP stack:

Octoparse MCP — Structured journalist extraction (primary data source)

Supabase MCP — Schema management, RLS policies, Edge Function deployment

GitHub MCP — Automated commits across the GlafyCo org

GlobalProxyManager — Custom geo-routing layer for multi-region journalist discovery across 100+ geographic IPs

The combination of Octoparse (extraction) + Claude (reasoning) + Supabase (persistence) creates a closed-loop intelligence system where every journalist match is grounded in real, live web data.

Production Philosophy — Pipeline Not Repository
One design decision worth sharing: Octoparse MCP is never used as a data warehouse.

Every extraction is immediately scored, matched, and actioned. Data follows a strict TTL policy:

Days 1–30: Hot storage — full access, edit, download

Days 31–60: Cold storage — read-only, raw source stripped

Day 61: Hard delete

This keeps infrastructure lean and reinforces the product positioning: E_MediaScience is a processing engine, not a data repository. Users ingest, score, pitch, and clear the decks.

Pricing Model
E_MediaScience uses a Core + Engines modular pricing architecture:

Core Platform — $69/month (dashboard, 2 seats, campaign management)

Signal Engine bolt-on (EMS) — from $29/month

Production Engine bolt-on (Clipositing video engine) — from $29/month

Agency tiers — from $799/month with HighLevel CRM integration

No credits. No per-minute charges. No "credit hostage-taking." Flat session-based pricing that scales by tier, not by usage clock.

Repo
Everything is open and committed:

GitHub: github.com/GlafyCo/E_MediaScience

The architecture docs, sprint plans, and HungQueryResolver spec are all in docs/strategy/. The multi-tenant core, AI scoring engine, and Supabase migrations are all there.

Built with Octoparse MCP + Cursor + Claude for the Octoparse MCP Challenge 2026.

Ian Taylor — Founder, GlafyCo | Wales, UK
Building E_MediaScience, Clipositing, and the GlafyCo AI platform stack | X: @ianbuildsagents