nodecode-venice-web

Venice AI web search and page extract: privacy-first search and Markdown page fetch, on a Venice API key. A Nodecode rewrite of MnrGreg/hermes-venice-web by MnrGreg.

The model gets two tools: venice_search, which returns titles, links and snippets for a question, and venice_extract, which fetches each URL as Markdown. Both talk to Venice's own /augment endpoints, which route a search to Brave (zero data retention) or Google (anonymized). /venice QUERY is the operator's own search, and (help :venice) is the model's page.

Get it

Tell the agent "add the venice-web cell", or run nodecode add nc://venice-web.

Set it up

A Venice API key from venice.ai, in the environment:

echo 'VENICE_API_KEY=…' >> ~/.nodecode/.env

or in the cell's section, which /settings writes:

{
  "venice": {
    "api_key_env": "VENICE_API_KEY",  // or "api_key" written inline
    "search_provider": "brave"        // brave (zero data retention) or google (anonymized)
  }
}

/venice with no argument says where the cell points. Each search and each scraped URL costs about $0.01 on that Venice account; the cell says so in every answer's header and in the tool descriptions, and max_urls caps what one venice_extract call may spend.

Setting	Default	What it does
`api_key` / `api_key_env`	none	the key, inline or by environment variable (`VENICE_API_KEY` is read either way)
`base_url`	`https://api.venice.ai/api/v1/augment`	the augment root, for a proxy of your own
`search_provider`	`brave`	the engine Venice searches with
`max_results`	`10`	results one search asks for, at most 20
`max_urls`	`5`	URLs one extract fetches, each billed
`timeout_seconds`	`20`	one search's ceiling
`scrape_timeout_seconds`	`30`	one page fetch's ceiling

Use it

venice_search — {query, limit?, provider?}. Answers one line per result: its title, its URL and a snippet clipped to 400 characters. limit is clamped to 1..20, provider to brave or google; unset, the section decides.
venice_extract — {urls}. One request per URL, in the order given, each page's Markdown under its title. A URL that fails says so on its own line and the rest still answer, as the original's per-URL error objects do.
/venice QUERY — the same search, for the operator.

What is in it

File	What it is
`cell.lisp`	the section, the two endpoints, the two tools, `/venice` and `(help :venice)`
`test/cell-test.lisp`	lifecycle, both tools, `/venice` and the refusals, every request a stubbed `nlk:http`
`config.example.jsonc`	the section, commented

Why two tools, not a search backend

In Hermes the plugin registered itself as the harness's web search provider, behind the standard web_search and web_extract tools. Nodecode's nodecode-websearch cell keeps its engines in a fixed list a cell cannot add one to, so Venice is two verbs of its own here. The endpoints, the request shapes and the answers are the original's; only the seam differs.

Test

From the cells repository's root, offline:

.github/test-cell.sh tools/nodecode-venice-web

nc://venice-web