nodecode-parallel-search
Parallel Search: keyless web search and page fetch through Parallel's free MCP server, as two model tools. A Nodecode rewrite of parallel-web/search-mcp by parallel-web.
Get it
Tell the agent "add the parallel-search cell", or run nodecode add nc://parallel-search.
Set it up
Nothing is required. The endpoint is anonymous: no account, no API key, no environment variable. The cell registers two tools, web_search and web_fetch, and answers a manual topic, /parallel for the operator.
{ "parallel-search": { "max_results": 5, "timeout_seconds": 60 } }
| Member | Default | Notes |
|---|---|---|
endpoint |
https://search.parallel.ai/mcp |
the MCP server's address; point it at your own if it speaks the same two tools |
max_results |
5 |
what a search asks for when the call gives no count |
timeout_seconds |
60 |
seconds one search or one page read may take |
Then (restart-cells "nodecode-parallel-search"), or set it in /settings. Say /parallel fastapi response model to try it.
The free tier is rate limited (HTTP 429) and meant for personal agents; an authenticated Parallel account raises the limits, and the cell says so when the limit is hit. If another cell already registers web_search and web_fetch, keep one of them.
Use it
Ask for anything current: official documentation, a release note, a fact you would otherwise guess. The model calls web_search and gets ranked results with excerpts and their urls; it cites them and fetches a page only when the question needs the page itself.
From eval, the same two calls:
(parallel:search "fastapi response model") ; one query
(parallel:search (list "fastapi response model" "pydantic v2 model_config")
:objective "what the latest FastAPI docs say")
(parallel:fetch "https://modelcontextprotocol.io/introduction")
(help :parallel) answers the same guidance the model reads; *primer* is the text, so a layer can reword it.
