Web Search
Let the model search the web, read the pages it finds, and cite them in its answer.
Web search gives the model a web_search tool. When a question needs current information, the model writes a query, LibreChat fetches results, reads the top pages, and the answer comes back with the sources it used.
Turn it on in a chat
Click + in the message composer to open the palette, then pick Web Search in the Tools section. It shows up as a chip on the composer while it is on; remove the chip, or pick it again in the palette, to turn it off. You can pin it so the chip stays on the composer even when it is off.
With web search on, the model still decides for itself whether a message needs a search. Asking "what changed in the latest Node.js release?" will usually trigger one; asking it to rewrite a paragraph will not.
For agents, web search is a capability you enable in the agent builder. Anyone chatting with that agent gets the tool without toggling anything.
What you see in the answer
While the tool runs, the message shows its progress ("Searching the web", then "Reading results"). Once it finishes, it collapses to Searched the web; expand it to see the pages it found, or use the Search details button next to it to see the query the model sent.
Inline citations in the text point back to those pages, so you can check a claim against the page it came from. When several agents answer side by side, a Sources row above the replies also collects the results, with All, plus News and Images tabs when the search returned those.

How a search runs
Each search goes through up to three services:
- A search provider returns a list of results: Serper, SearXNG, Tavily, or Keenable.
- A scraper fetches the pages behind those results and extracts their text: Firecrawl, Serper Scrape, Tavily Extract, or Keenable Fetch.
- A reranker (Jina or Cohere) picks the passages most relevant to the query. Reranking can be set to None to skip this step.
Only the extracted passages reach the model, not whole pages. Pages that time out or block scraping are skipped, so an answer can cite fewer sources than the search returned.
Safe search is set to moderate by default. Your admin can raise or lower the level for everyone; Tavily is the exception and only filters when the admin turns on its own safe search option.
Adding your own API keys
If your admin has not supplied keys for every service, turning on Web Search opens a dialog asking for them. It has a section for each service the admin left to users, out of Search Provider, Scraper, and Reranker. Pick a service in each, fill in the keys (each field links to where you get one), and click Save. SearXNG asks for your instance URL, including http:// or https://, such as https://searxng.example.com, and its key is optional. See URL Formats for what each custom URL field expects. Keenable works without a key; adding one only raises its rate limits.

The keys are stored with your account and only used for your searches. To change provider or replace a key later, open the palette and click Configure (the gear) on the Web Search row. Once you have saved keys, Revoke in the same dialog deletes them. When the admin has set every key on the server, there is no dialog and no gear.
A URL you enter is never paired with a key the admin set on the server. If the admin set a Firecrawl or SearXNG key and you enter your own URL for that service, enter your own key too. Otherwise Firecrawl ignores your URL and uses its default endpoint, and SearXNG does not run. A keyless SearXNG instance still works: when no key is involved, your URL alone is enough.
Web Search option is not visible
There is no separate web search button; the option only appears in the palette's Tools section. It is missing when:
- You are chatting with an agent, or on the Assistants endpoint. Agents carry their own tools, so web search is turned on in the Agent Builder instead.
- The selected model spec hides the tools row.
- Your role does not have the Web Search permission. That also blocks the built-in search some providers offer.
- Your admin removed
web_searchfrom the agent capabilities.
Setup
Web search is available by default for agents and chats, but it needs a search provider and scraper before it returns anything. An admin either sets the keys on the server or leaves them for users to enter. Self-hosted providers on a private address are blocked until the admin lists them in allowedAddresses. See Web Search configuration for providers, keys, and options, and Setting up SearXNG to self-host the search provider.
How is this guide?