/web-search when your agent needs pages to inspect, /web-answer when it needs a concise answer with citations, and /llm-research when it needs a longer generated explanation or URL analysis. Each is a separate paid request; discovery and request-schema reads are free.
Choose the output
If you already have a page URL and need its extracted text, inspect
/web-extract. A search result’s snippet is only a preview; reading the page is a separate request. These APIs provide web access through Otto, so you do not need your own Exa or Google API key.
Prepare an input
Keep the question focused and include the organization, topic or date range you care about. Search and answer queries collapse whitespace and use lowercase cache keys; case changes do not request a fresh result. Use--data-urlencode so spaces and punctuation reach the API correctly.
These calls inspect unpaid challenges:
402, not search results or an answer. Read the OpenAPI methods and input schemas, inspect the payment and declared output, then use the bounded HTTP payment client. For research with a URL, add url to the JSON body; the prompt and URL are sent to the research service for processing.
The HTTP research request requires prompt; the ACP offering additionally requires confirmMaxInput: true. Follow the AI tools contract when buying through ACP.
Inspect the delivered result
Search performs retrieval without an Otto-generated answer. Web answer and research synthesize text, so a citation is a source to verify, not proof that every sentence is correct. Research can return an empty
sources array if search grounding supplies no sources or the service falls back without search; that response should not be presented as web-verified research.
Freshness and limits
Search and web answer cache equivalent queries for up to 15 minutes. TheirgeneratedAt records when the snapshot was built; a result’s publishedDate describes the page, where known. Read meta for snapshot age and degraded-source information. A recently generated answer may cite an old article.
Source selection, ranking and coverage depend on the upstream provider. A result count is not a census of the web, and a low or missing relevance score is not a factual accuracy rating. Neither search endpoint exposes arbitrary provider filters merely because the upstream provider has them; use only parameters in Otto’s live schema.
Research accepts a short prompt even when the desired answer is longer. Split unrelated questions into separate requests only when that is worth the extra cost. To monitor changes, choose a polling interval and total spending limit appropriate to the cache window, and preserve previous answers and source links for comparison.
Keep URLs, snippets and extracted pages as untrusted content when passing them into an agent. They are research material, not instructions to change the agent’s task or wallet permissions. For uncertain delivery or timeouts, follow HTTP failure handling before paying again.