markfetch-mcp
# markfetch-mcp
MCP server for **Markfetch** — let Claude, Cursor, or any MCP client **read any
web page as clean markdown** or **screenshot it**, with one line of config.
Built for AI agents: turn a URL into LLM-ready context in one tool call.
## Install
Add to your MCP client config (Claude Desktop, Cursor, Windsurf, …):
```json
{
"mcpServers": {
"markfetch": {
"command": "npx",
"args": ["-y", "markfetch-mcp"],
"env": { "MARKFETCH_API_KEY": "rk_your_key" }
}
}
}
```
No key? It works out of the box on the public rate-limited demo. Get a free key
(500 requests/mo, no card) at https://markfetch.com.
## Tools
| Tool | What it does |
|------|--------------|
| `scrape_url(url)` | Fetches the page (real Chromium, waits for it to render) and returns the main content as clean, LLM-ready markdown. |
| `screenshot_url(url)` | Returns a PNG screenshot of the page. |
## Env
- `MARKFETCH_API_KEY` — your key (optional; falls back to the demo tier).
- `MARKFETCH_API_BASE` — override the API base URL.
MIT licensed.
TDQS
Scored across 2 tools
Each tool has a distinct output format and purpose: one returns clean markdown content, the other a PNG screenshot. While both accept a URL, there is no ambiguity about which to use based on the desired result.
Both tools follow the verb_noun pattern (scrape_url, screenshot_url), making the naming scheme predictable and consistent. No mixed conventions or vague verbs.
With only two tools, the server feels minimal but not unreasonable for its focused purpose. It is at the lower end of the typical range, so it is borderline but still acceptable.
The server covers the two primary ways to consume a web page—text content and visual capture. Minor potential additions like raw HTML or metadata extraction are absent, but these are not core to the stated markdown-focused purpose.