Web search in the chat

The chat can call a web search tool so the model can look up current information. Hosted search is built in and billed per search (opt-in); if you run your own search MCP server, you can connect it in chat settings instead.

Hosted web search (Linkup, per-search fee)

  • Open the capabilities menu (the sliders button next to the plus in the composer) and tick Web search (off by default). The model then gets a web_search tool; each search bills a flat fee from your prepaid credits (the menu shows the current price).
  • Queries are handled by Linkup, a search API subprocessor. The query travels from your browser to ozeye’s EU infrastructure and then to Linkup, which processes it in the EU under zero data retention: the query is used in memory to produce results and is never stored.
  • Results are returned to your browser and are never stored by ozeye.

Search modes

The capabilities menu decides which hosted tools the model has; the model still decides when to use them, after reading your message. Nothing runs before the model turn. Choices are remembered per chat and become the default for new chats. Every capability is always listed; when the selected model cannot use one (no tool support, no adjustable thinking) it is greyed out with the reason.

  • Web search off: no search tools are attached; the model answers from its own knowledge.
  • Web search on: the model gets web_search and fetch_page with strict guidance to search only when the answer depends on recent events or specific external facts, and a hard budget per answer: 2 searches and 1 page fetch by default, or 5 or 10 searches (with about one page fetch per two searches) when you raise it under Searches per answer in the capabilities menu.
  • Research (applies to the next message): adds a web_research tool that runs one Linkup deep search (multi-step search-and-read, 5 to 30 seconds, billed at the deep-search price). The model forms the query itself, uses it once, then answers with citations; budget 1 research call, 2 searches, 2 fetches.
  • When a budget is spent, the tool returns a note asking the model to answer from what it has, so an answer can never spiral into an open-ended search loop.

Privacy implications of hosted search

  • Your query text leaves ozeye’s servers: it is sent to Linkup, which processes it under its own privacy policy (EU processing, zero data retention, no training).
  • ozeye stores only billing metadata: user, API key, search modality, cost, and timestamp. The query text and the results are never logged, cached, or persisted (ZDR). Audit records contain only the query length, not the query.
  • Page fetches (fetch_page) do not go through Linkup: ozeye’s server requests the page from the website directly, so the site sees ozeye’s server, not you. The URL and the page are not stored.
  • Linkup is listed as a subprocessor; see the subprocessors page.
  • A self-hosted search MCP server (your own infrastructure) never touches ozeye servers beyond the same-origin proxy you explicitly enable.

Zero Data Retention posture

The ZDR promise covers content that passes through ozeye infrastructure. Hosted search passes only the query (never stored) and billing metadata (stored); results never touch our servers beyond being relayed to your browser.

Hosted search is opt-in and off by default: nothing is ever sent to Linkup unless you enable the setting.

Bringing your own search server (optional)

If you run your own search MCP server (any MCP server exposing a search tool), you can connect it in chat settings → MCP. The ozeye same-origin proxy forwards your browser’s requests to your instance. In production it blocks private/loopback addresses (SSRF guard), so your instance needs a public address, e.g. a subdomain behind your own reverse proxy with TLS. For local development, the proxy can allow loopback/private targets with OZEYE_MCP_PROXY_ALLOW_LOCAL=1. The server typically expects an Authorization: Bearer token of your choice; the proxy forwards it to your instance.

Billing

Hosted search bills a flat per-search fee from your prepaid credits; a Research deep search bills the separate deep-search fee shown on the pricing page. Unsuccessful searches (e.g. the search backend is unavailable) are not billed. A self-hosted search server is free to use: it runs on your infrastructure.