nc://ollama-search

Ollama Cloud Web Search: web search and page fetch through Ollama Cloud's native API, on the Ollama API key a Cloud profile already has. A Nodecode rewrite of rriggs/hermes-plugin-ollama-cloud-search by rriggs.

The original is a web search provider for a harness's own search backend: it advertises search and content extraction, and answers both by asking two Ollama endpoints, POST /api/web_search and POST /api/web_fetch, with OLLAMA_API_KEY as a bearer. Nodecode's search cell keeps its engines in a fixed list a cell cannot add one to, so this cell offers the same two calls as two tools of its own, and /ollama-search for the operator.

Get it

Tell the agent "add the ollama-search cell", or run:

nodecode add nc://ollama-search

Set it up

One key, and it is the one Ollama Cloud chat completions already use: a key from https://ollama.com/settings/keys, in OLLAMA_API_KEY. A profile serving Ollama Cloud models needs nothing else.

{
  "ollama-search": {
    // the variable the key is read from; or write the key here as "api_key"
    "api_key_env": "OLLAMA_API_KEY",
    // the host the two endpoints are asked on
    "base_url": "https://ollama.com",
    "max_results": 5,
    "timeout_seconds": 30,
    "fetch_timeout_seconds": 60
  }
}

A section that names a variable with nothing in it refuses at start, and the refusal says which variable and where a key comes from. base_url moves every call to another host, for a proxy or a compatible server.

Use it

The model has two tools while the cell runs:

• ollama_search {query, limit?} — titles, urls and a snippet per result, at most 10, as Ollama's search returns them.
• ollama_fetch {url} — one page as the text Ollama fetched and extracted from it.

(help :ollama-search) is the primer behind them. From eval, the same calls are (ollama-search:search "query" :limit 5) and (ollama-search:fetch "https://…"), each answering a string. /ollama-search QUERY searches from the operator's side; with no argument it says which host and how many results a search asks for.

What changed

• The endpoints, their request bodies and their answer shapes are the original's: /api/web_search with a query and a count capped at 10, /api/web_fetch with one url per call.
• The original also let OLLAMA_BASE_URL move its calls; here that is the section's base_url, so the operator sets it in one place.
• Its "free tier" badge and its setup-panel schema are the harness's onboarding, not a capability: the section and its refusal are this cell's.
• It honoured the harness's interrupt flag between calls; a Nodecode turn is cancelled by its own condition, which passes through the tools untouched.

Test

From this repository's root:

.github/test-cell.sh tools/nodecode-ollama-search

Every request is a stubbed NLK:HTTP and the key a stubbed NLE::CREDENTIAL-ENV, so nothing reaches Ollama or the network.

MIT licensed.