nodecode-crawl4ai
Crawl4AI Web Extract: turns any page into clean markdown through your own self-hosted Crawl4AI instance. A Nodecode rewrite of mblauser/hermes-plugin-crawl4ai by mblauser.
Crawl4AI is an open-source crawler that renders a page in its own browser and answers it as LLM-friendly markdown, so a JavaScript-heavy site extracts as readily as a static one. It runs on your machine or your own server: no page, and no text, leaves your network. The server is yours to run; the cell never installs, starts or stops it.
Get it
Tell the agent "add the crawl4ai cell", or run nodecode add nc://crawl4ai.
Set it up
Run the server first:
docker run -d -p 11235:11235 unclecode/crawl4ai
It prints a JWT on startup when auth is on. Then give the cell its address in /settings, or in ~/.nodecode/config.jsonc:
{
"crawl4ai": { "url": "http://localhost:11235", "api_token": "the-jwt-it-printed" }
}
url is required and must be http:// or https://; api_token is optional, for a server started with auth off. timeout_seconds (default 60) is how long one page may take — the server renders it in a browser first. config.example.jsonc in this folder lists every member. A url that is not an http address refuses the start, and the folder's line in /cells says why.
The route it calls is the one Crawl4AI serves: POST /md with {"url": ...} and a bearer token. A server behind a path prefix works — put the whole prefix in url, without a trailing slash.
Use it
Ask the agent to read a page, or call the tool yourself:
| Tool | What it does |
|---|---|
crawl4ai_extract |
One or more page addresses in, each page's title, address and markdown out — or one ERROR line for the page that failed. |
/crawl https://example.com extracts one page without asking the model, and typed alone it asks for the address.
A page address pointing into a private, loopback or link-local network is refused before the request goes out, as the original plugin's SSRF check refused one. The bearer token never appears in an answer: every refusal is filtered for it.
The cell answers the same three shapes the original read: a markdown string, a markdown object carrying raw_markdown, and the title taken from the page's first # heading. A failed page is answered as ERROR: with the status and the server's own detail, so the model can repair the call instead of repeating it.
To switch it off, drop the folder or set "crawl4ai": {"enabled": false}: nothing is registered then.
