Web Search and Host Intelligence

Two toolkit types let an agent reach the public internet. Both run on the Rust worker and are read-only.

Web Search (web): web_search and web_fetch. web_fetch reads a page through the platform’s egress guard, which refuses private and metadata addresses and checks every redirect, or through a provider’s extract API.
Host Intelligence (host_intel): internet_host_search and internet_host_lookup, through Shodan. It is off until a platform administrator grants it to a project in Admin › Toolkits.

Every call is admitted by the same budget engine as model calls before the provider is called, and billed afterwards. Tool spend appears in Settings › Usage as rows such as “Web search · Tavily”, next to the model rows, and in run analytics. When a project or member budget runs out, the tool is refused with the same message and Usage link as a refused model call, and the provider isn’t called.

Set up a provider

Enable the provider
In Admin › Configuration › Web search, add the provider to Enabled providers. A configured key alone doesn’t turn a provider on. The same section sets the default search provider, whether direct fetch is allowed, the fetch allow and deny domain lists, and the result caps.
Add a credential
Create a tavily, firecrawl, brave_search, duckduckgo_instant or shodan credential. A project uses its own credential first. If it has none, it uses a shared credential in the public project, except for Shodan, which only accepts a project’s own key.
Check the price
Open Admin › LLM proxy › Tools. Each operation has a price per request, credit or page. An operation without a price is refused, and free operations have an explicit $0. Change the price if your plan differs from the default.
Attach the toolkit
Create a Web Search or Host Intelligence toolkit and attach it to an agent. On the desktop, the native client policy must also allow it (local_work.web, local_work.host_intel, both off by default).

Providers

Prices, limits and terms change, so check them with the provider before you rely on them. The default prices are the ones the Tools pricing tab starts with.

	Tavily	Firecrawl	Brave Search	DuckDuckGo Instant Answers	Shodan
Toolkit	Web Search: search and fetch	Web Search: search and fetch	Web Search: search	Web Search: search (instant answers only)	Host Intelligence
Pricing	Credits: basic search 1, advanced 2, extract 1 per 5 URLs. Pay as you go $0.008 per credit	Credits: 1 per scraped page, 2 per 10 search results. Hobby $19/month for 5,000 credits, up to Scale $599/month for 1M	$5 per 1,000 search requests	Free	Query credits: Freelancer $69/month for 10,000 (about $0.0069 each), Small Business $359/month for 200,000
Default price	$0.008 per credit	$0.0038 per credit or page	$0.005 per request	$0	$0.0069 per credit
Free tier	1,000 credits a month, no card	1,000 credits a month, no card	$5 of credits a month (about 1,000 searches), card required	Free, no published quota	A free API plan, but most methods need at least Membership ($49 once). Academic accounts are non-commercial only
Rate limits	100 requests a minute on a development key, 1,000 on production	Free: 10 a minute for search and scrape. Standard: 500 a minute	50 queries a second on the Search plan	Not published	1 request a second on every plan
Storing results	Terms don’t cover caching results. Tavily keeps a perpetual licence to your queries and URLs	Terms don’t cover it. Firecrawl caches pages itself (2 days by default). Zero data retention needs Enterprise	Transient storage only. No database of results	DuckDuckGo says it can’t fully syndicate its results	Not covered. No redistribution of the content
Attribution	Not required	Not required	Optional. If shown, “POWERED BY BRAVE” with the logo	Credit the source (usually Wikipedia)	Required. Every result carries “Data from Shodan (shodan.io)”
AI use	Allowed. Tavily may train on its output. No competing models, no resale	Allowed. You’re responsible for lawful use of what you scrape	Grounding is allowed. No training or fine-tuning on results	Not covered	Not covered
Lifecycle	Active	Active	Active	Active. It isn’t a full search API, so it’s never the default provider	Active
Shared key allowed	Yes	Yes	Yes	No key needed	No. The terms bar reselling or redistributing one account’s data, so each project uses its own key

Elitea never stores results beyond the turn. Logs and audit record hashes, hosts and looked-up addresses, never full URLs or queries.

Available tools

Web Search (2)

Tool	What it does
`web_search`	Searches the web through the toolkit’s default provider, or the platform default
`web_fetch`	Reads one page as Markdown or text, directly or through a provider

Host Intelligence (2)

Tool	What it does
`internet_host_search`	Searches Shodan’s index of internet hosts. Raw banners are never returned
`internet_host_lookup`	Looks up one public IP address. Private addresses are refused without a provider call