v1.1 Now with Readability · cleaner content, sharper reads

Your AI keeps saying "I can't access that link."

Millions of AI users hit this wall every day. McpFOX fixes it from the other side: your browser becomes your AI's eyes. Real Chrome, real TLS, real cookies — running locally, nothing uploaded, free forever for individuals.

★★★★★ Chrome Web Store 4.8 50,000+ users MCP 2025-03-26 MIT open core
SUCCESS
91.4%
LATENCY
0.82s
🔒 nytimes.com/section/technology
// photo placeholder //
C
Claude Desktop
via fox_fetch

The core argument of the piece:

The real bottleneck for AI agents isn't model quality — it's whether they can reach the live, real web at all.

The article lays out three case studies

fox_fetch ✓ fox_batch fox_search
Chrome Store4.8/5
Active users50,000+
Avg response<1s
Success rate91%
Zero uploads100% local
LicenseMIT · OSS
Chapter one · Why you're here

Three ways the web
quietly ruins your AI workflow.

Models keep getting smarter. Their eyes are still stuck in 2023. The real problem isn't that AI is dumb — it's that an invisible wall keeps locking it out of the live web.

"I'm unable to access that link."

You paste a URL. Back comes a hedged memory from 2023 training data — not the article that exists right now. The exact page is one click away for you, and a wall for your AI.

> Summarize this article:
  https://example.com/ai-breakthrough

Claude: I'm sorry, I can't
directly access that link. Based
on my training data, I can
infer that…

$300+ a month of invisible tax

Bright Data, ScrapingBee, ZenRows — scraping services aren't cheap. Worse, every request routes through someone else's servers. Your research walks out the back door with it.

Bright Data         $199
ScrapingBee         $ 49
Proxy pool          $ 79
─────────────
TOTAL             $327 /mo

Copy. Paste. Again. Again.

Open page, select all, copy, switch apps, paste, wait, next URL, repeat. A five-source research session burns thirty minutes on pure logistics — before a single insight lands.

⌘A⌘C⌘⇥⌘Vwait…⌘A⌘C⌘⇥⌘Vwait…again?
Chapter two · Another way

Make Chrome itself
your AI's second pair of eyes.

Not a scraper. Not a proxy. Not a cloud API. McpFOX makes the browser the bridge — real TLS fingerprint, cookies you're already signed into, local parsing. Your AI reaches it over MCP, and nothing leaves your machine.

01 — Source

Your own Chrome

Real user fingerprint, full HTTP/2 stack, the login state you already have. The target site sees a real visitor — not a suspicious bot.

Native TLSCookies inherited
02 — McpFOX · processing

Local parsing & structuring

Readability strips the article clean — no ads, nav or boilerplate. Output as Markdown, JSON, HTML, text or PDF. Everything happens on your device.

Zero uploadsL1+L2 cache5 formats
03 — Endpoint

Any MCP-speaking client

Claude, Cursor, Cline, Continue, ChatGPT Desktop — plain JSON-RPC 2.0 over MCP. Configure once, wired forever.

MCP 2025-03-266+ clients
Chapter three · Anatomy

Six small things,
one well-sharpened chisel.

Every feature answers a single question: fetch succeeds, latency stays low, format fits the model, privacy holds, integration is trivial, extension is possible. We didn't compromise on any of them.

F—01 · Anti-bot penetration

A 91% success rate.
We don't "pretend" to be a browser. We are one.

Real Chrome TLS fingerprint, HTTP/2 frame ordering, rotating headers, exponential backoff. Cloudflare, PerimeterX and Akamai are all effectively transparent to it.

91%
F—02 · MCP native

One block of JSON.
Your AI just knows.

// claude_desktop_config.json
{
  "mcpServers": {
    "mcpfox": {
      "url": "http://localhost:7431"
    }
  }
}
F—03 · Clean content

Powered by Readability.
Only the article survives.

The same algorithm behind Firefox Reader and Pocket. Ads, comments, related-posts rails — stripped before the model ever sees them.

F—04 · Format outputs

Five formats, cut to the model's appetite.

Token spend drops 60% on average. Structured JSON for agents. Markdown for chat. Plain text for embeddings. HTML when you need to debug.

MDMarkdown · the default for AI chat1.2k tok
JSONStructured · for agents and pipelines0.9k tok
HTMLRaw · for debug and reprocessing4.7k tok
TXTPlain text · for embedding indexes1.0k tok
PDFDocument · for long-form extractionauto
F—05 · Local-first

Your content
never leaves your machine.

Two layers: in-memory L1 and IndexedDB L2 — both on-device. Fetch results never touch our servers or a third party. This is the default, not a paid upgrade.

F—06 · Parallel batch

Fifty URLs,
all in flight.

A built-in semaphore and per-domain rate limit keep you below every anti-bot threshold.

Chapter four · Getting started

Three steps,
and your AI actually reads the web.

Under five minutes end-to-end. No account. No credit card. No README required.

i.

Install the extension

Open the Chrome Web Store and hit "Add to Chrome." A 🦊 appears in your toolbar — you're ready.

~30 seconds · no signup
$ Extension installed ·
ID: gcjkfpldmbejjmmoa...
Runtime: McpFOX @ localhost:7431
ii.

Configure your AI

Open the config file in Claude Desktop or Cursor. Paste the JSON we generated for you. Save, restart. Done.

~2 minutes · once, forever
"mcpServers": {
  "mcpfox": { url: "…:7431" }
}
iii.

Start asking

Say "read this article for me." McpFOX quietly calls fox_fetch — fetch, clean, parse, return. You only see the answer.

Unlimited (Free: 200/day)
You: read example.com/post
Claude: [calling fox_fetch...]
The three key points of this piece…
Chapter five · Live test

Don't take our word.
Run both methods, right now.

Paste any URL. See what a traditional scraper gets back versus what browser-native McpFOX gets back. A simulation for now — after install, the right side becomes your daily reality.

Same URL. Two worlds.

Left: life without McpFOX. Right: life with it. Running simultaneously.
Try →
medium.com
techcrunch.com
arxiv.org
twitter.com
news.ycombinator
Without McpFOX · traditional scraper
Idle
Waiting to run…
Via McpFOX · browser-native
Idle
Waiting to run…
Chapter six · Pricing

Start free.
Upgrade when you need to.

The core fetch capability is free forever for everyone. Paid tiers solve scale and team problems — never the core itself.

Free
$0/forever
No credit card · personal use
  • 200 fetches per day
  • All MCP tools
  • All 5 output formats
  • Baseline anti-bot
  • L1/L2 local cache
  • Rendered-mode fetch
  • Cloud history sync
Install free
Team
$99/month
Up to 10 seats · SSO
  • Everything in Pro
  • Shared domain rule library
  • Team usage dashboard
  • Webhook integrations
  • SSO single sign-on
  • 4h priority support
Contact us
Enterprise
Private deployment · compliance audit trails · SLA · custom rules · dedicated CSM · starting at $399/month
Talk to sales →
Chapter seven · Early users

Some of them felt their AI truly go online
for the first time that afternoon.

"
I spent three months wrangling proxy pools and scraper stacks. The thing I actually needed was for Claude to just read web pages directly. The moment McpFOX installed, I deleted all three scraping pipelines we had.
M
@matchalatte_dev
AI Agent Engineer · Stripe
"
I'm not a programmer, but I need Claude to read forty industry reports a day. Copy-paste used to eat two hours. Now it's one sentence and thirty seconds.
L
Lena C.
Industry analyst · Tokyo
"
Using the real browser as the MCP source is the right direction. Privacy, compliance and penetration — all three solved in a single move. Our entire agent team switched.
K
Kei Tanaka
CTO · Autonomous Labs
Chapter eight · Questions

The things people ask first.

Our eight most-asked questions. See the full FAQ → or come poke us on Discord.

Really free. All five MCP tools, all five output formats, nothing paywalled. The only quota is 200 fetches per day — roughly one every seven minutes if you spread them evenly. Enough for almost every individual. We refuse to ship a crippled free tier as a pressure tactic.
No. Local-first is a design principle here, not a marketing line. Fetch results live only in L1 memory and L2 IndexedDB — both on your device. There is no "McpFOX relay" server; we don't operate one. You can unplug from the internet and still hit previously cached pages.
Three things break in the cloud. Server IPs get identified and blocked. Your queries flow through our pipe. The service requires a subscription. Local browser fetching kills all three in a single move — that's the whole idea.
Native support for Claude Desktop, Cursor, Cline, Continue, ChatGPT Desktop, and any client that speaks MCP 2025-03-26. Windsurf and Zed integrations are on the way.
No. Configuring your AI is four steps: copy, paste, save, restart. We generate the config for you at install time — you just drop it in the right folder.
Paid plans activate immediately on checkout. Unhappy for any reason within 14 days? Email billing@mcpfox.space — we refund in full. No questions asked.
The core fetch engine is MIT-licensed. Commercial pieces — cloud sync, team dashboards, enterprise integrations — stay proprietary. Open core plus commercial periphery is, in our view, the honest sustainable model.
Private deployment, SOC 2 compliance audit trails, SLA guarantees, a custom domain-rules service, a dedicated customer success manager, priority fetch lane and onboarding training. Built for regulated environments and larger teams.

Let your AI really
read the web.

Install for Chrome — free ⭐ Star on GitHub
MIT · Fully local · Free forever for individuals · 2-minute setup

Tweaks ×

· Live preview
· Close with × when done