The agent has two tools that reach the public web: web_search, which
searches, and web_extract, which fetches a specific page. Both are sensitive
calls, so they ask for approval in Ask and Auto.
Pages the agent fetches are added to the conversation as sources, so it can cite them and you can read what it read.
Two ways to search
Your model provider's built-in search. Anthropic and OpenAI models can run a search on the provider's own servers as part of the turn. Nothing leaves your machine except the request you were already making, and there is no extra key to configure. Gemini models do not offer this in Tidebreak today.
A search provider you configure. Tidebreak calls the search API directly from your machine.
| Provider | Notes |
|---|---|
| Exa | Search and page extraction. |
| Tavily | Search and page extraction. |
| Brave Search | Search only. Has a free tier. |
| SearXNG | Self-hosted. No key — you provide the instance URL. |
When the chosen provider cannot extract a page, web_extract falls back to
Tidebreak's own fetch-and-read path rather than failing.
Configuring it
Settings → Web search. Search mode decides which of the two paths is used:
| Mode | Behavior |
|---|---|
| Automatic | Use your configured provider if there is one; otherwise the model provider's built-in search; otherwise nothing. |
| Model provider (built-in) | Always the provider's own search. Turns off on models that do not support it, rather than falling back. |
| Configured provider | Always your own key. |
Below that, save keys for as many providers as you like, then pick the Active provider. Disabled turns search off entirely. There is also a request timeout (default 20 seconds).
Saved keys are never displayed back to you — the panel shows Ready or Not configured per provider.
On a fresh install with no search key, Automatic gives you working search on Anthropic and OpenAI models through the provider's own search, and nothing on Gemini models until you add a key.