# cleanfetch

An **MCP server** that gives AI agents clean, reliable web reading. One job, done well.

| Tool | Input | Output |
| --- | --- | --- |
| `fetch_clean` | a URL | the page's main content as clean **Markdown** + title/byline (nav/ads/scripts stripped) |
| `extract_links` | a URL | all hyperlinks with anchor text, classified internal/external |
| `extract_metadata` | a URL | title, description, canonical, robots, **Open Graph**, **JSON-LD**, headings outline |

## Why this exists

Agent builders constantly need to *read web pages* — but the generic fetchers choke on
real-world HTML, return nav/ad noise, or get blocked. **Reliability is the moat**: a server
that consistently returns clean, structured content is worth paying for. There are 14,000+
MCP servers listed and almost none are production-hardened or monetized — that's the opening.

## Run it locally

```bash
npm install
npm run build
npm start          # runs on stdio
# or, no build step:
npm run dev
```

Sanity-check the tools with the MCP Inspector:

```bash
npm run inspect
```

## Connect it to a client

**Claude Desktop / Claude Code / Cursor** — add to the MCP config:

```json
{
  "mcpServers": {
    "cleanfetch": {
      "command": "node",
      "args": ["/Users/alexandre/projects/mcp-cleanfetch/dist/index.js"]
    }
  }
}
```

Then ask the agent: *"Use fetch_clean to read https://example.com and summarize it."*

## Config (env vars)

- `CLEANFETCH_UA` — override the User-Agent.
- `CLEANFETCH_TIMEOUT_MS` — request timeout (default 15000).

---

## Go-to-market (the actual point)

**1. List it everywhere (free distribution).** Submit to the MCP directories — this is the
"App Store moment", and free-tier listings rank highest:
- PulseMCP — https://www.pulsemcp.com/
- Glama — https://glama.ai/mcp/servers
- Smithery — https://smithery.ai/
- Awesome MCP Servers (GitHub PR) — https://github.com/punkpeye/awesome-mcp-servers
- The official MCP registry.

Publish to npm so `npx mcp-cleanfetch` works, and add a 30-second demo GIF to the README.

**2. Harden the moat (what makes it paid-worthy).** Roadmap, in order of value:
- JS rendering fallback (Playwright) for client-side-rendered pages.
- Rotating proxies + retry/anti-bot handling for sites that block plain fetch.
- Caching + rate limiting.
- `batch_fetch` (many URLs in one call), `search` (query → top results → clean content).

**3. Monetize.** Free tier = local stdio (what's here). Paid tier = a **hosted HTTP MCP
endpoint** with an API key + **usage billing**:
- Host the server over the Streamable HTTP transport (the SDK supports it).
- Meter calls per key; bill with **Stripe** (Stripe shipped MCP/usage billing rails in 2026)
  or a usage-billing layer. Typical pricing in this category: a free tier + ~$20–40/mo, or
  per-call credits. Realistic target: **$0.5–3K MRR per server** — cheap enough to run several.

**4. Validate before hardening.** Ship the free version, list it, and watch installs/usage.
If a tool gets real pull, *that's* the signal to build the paid hosted tier. Don't build
billing for a server nobody installs.

## License

MIT