Skip to main content
Server tools run during the model call itself. Your app lists them in tools, the model calls them when it needs them, and the reply comes back with the answer already using their results. Your app never sees or runs those calls. Nemonist runs these tools itself, on every model and route: models from the catalog, your own provider keys, and custom endpoints alike. The model only sees each tool as a function it may call; nemonist does the search, the fetch or the clock reading and gives the model the result.

Tools

To make the model call one, name it in tool_choice: {"type": "nemonist:datetime"}. The model then calls it once and is free to answer. An app written for OpenRouter can keep its openrouter:web_search, openrouter:web_fetch and openrouter:datetime entries: nemonist reads them as the same three tools and runs them itself.

Cost and limits

  • The tools cost nothing on every plan. You pay only for the model’s own tokens, including the extra step in which it reads a tool’s result.
  • Daily limit: each account may run 100 web searches and 500 page fetches a day, counted across all its keys and models and reset at 00:00 UTC. The date tool has no daily limit. Past the limit, the tool tells the model it reached the daily limit, the model answers from what it has, and the request still succeeds (200). A refused call does not count.
  • In one reply the model can run up to 5 web searches, 5 page fetches and 10 date calls, unless the tool’s max_uses says otherwise. On the Free plan a reply may search at most 2 times and fetch at most 5 pages.

How they behave

  • nemonist:web_fetch reads public https pages only (an http URL is read over https). It refuses private and internal addresses, including after a redirect.
  • stop_server_tools_when is honored: after each step that ran a server tool, if any condition holds (step_count_is, max_tokens_used, max_cost, has_tool_call, finish_reason_is), the server tools are taken away and the model answers. Your own tools stay.
  • The reply shows each tool nemonist ran: on /responses, an output item whose type is the tool’s name as you sent it (status completed, with the fetched page or the search sources) before the answer; on /messages, a server_tool_use block with that name; on /chat/completions, only the usage counts below. You can send these items back in the conversation as they are.
  • A tool that fails (a page that does not load, a search that finds nothing) tells the model so, and the model answers from what it has. The request does not fail.
  • A model that ignores tool_choice may answer without calling the tool.
Provider tools you write in a provider’s own shape pass through to that provider as written: Anthropic’s web_search_20250305 on /messages, and {"type": "web_search"} on /responses.

Usage

Every server tool call shows in the reply’s usage, together with any calls made during memory’s own steps:

Unsupported tools

A server tool nemonist does not run (for example openrouter:code_interpreter) is refused with 400, and the message names it: