Loading…
One API key gives your agents live search results and the readable text of each page, over REST or MCP.
claude mcp add viberoo-searchAnimation: an AI agent asks a question, calls the web_search tool, receives three results with their page text, and answers with citations.
It calls the web_search tool over MCP, or POST /api/v1/search, with its workspace key.
Key, input and the per-key rate limit are checked before anything runs.
Your SearXNG finds results. With include_content, the top pages open in a sandboxed browser.
Titles, links, snippets and page text come back. The call is metered to the key.
Call it from your server with REST, or give it to your agent as an MCP tool. Same key, same results.
curl https://tools.viberoo.ai/api/v1/search \
-H "Authorization: Bearer $VIBEROO_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "postgres row level security",
"max_results": 3,
"include_content": true
}'postgresql.org/docs/current/ddl-rowsecurity.html
Tables can have row security policies that restrict, per user, which rows can be returned by normal queries…
+ main text of the page
postgresql.org/docs/current/sql-createpolicy.html
Define a new row level security policy for a table. A policy grants permission to select, insert, update or delete rows…
+ main text of the page
postgresql.org/docs/current/sql-altertable.html
These forms control the application of row security policies belonging to the table…
+ main text of the page
Agents need fresh facts with links they can quote. Viberoo AI returns both, in a shape a model can use without cleanup.
Set include_content and the top pages are opened in a sandboxed browser. You get each page's main text, up to 10,000 characters, next to its title and link. A page that can't be read says why; the search still succeeds.
{
"title": "Row Security Policies",
"url": "https://www.postgresql.org/docs/…",
"content": "In addition to the SQL-standard privilege
system, tables can have row security policies…",
"content_error": null
}Connect any MCP client to one URL and your agent gets a web_search tool. The tool tells the model that snippets and page text are third-party content, so it treats them as data rather than instructions.
{
"name": "web_search",
"arguments": {
"query": "postgres row level security",
"max_results": 3,
"include_content": true
}
}Bad input is refused before anything runs, with the field named. Over-limit keys get 429 with Retry-After, and every response carries a request id you can quote.
HTTP/1.1 400 Bad Request
{
"error": {
"code": "invalid_request",
"message": "max_results: Too big: expected number to be <=10"
}
}Anything that should check the web before it answers.
Answer questions with current sources, and quote the page text instead of guessing from a snippet.
Look up docs, changelogs and error messages from inside Claude Code, Cursor or your own agent.
Check vendor docs, status pages and release notes before answering a customer or a teammate.
Run searches from your server on a schedule and summarize what changed since last time.
Every rule below is enforced in server code, not in a prompt.
Only a SHA-256 hash of each key is kept. You see the key once, when you create it.
Page reading refuses loopback, private, link-local and cloud-metadata addresses, including through redirects.
Each search gets a new browser profile. Downloads, service workers and WebSockets are off.
Results come from a SearXNG instance you run on a private network, not from a third-party search API.
Three steps, about a minute.
Sign up with your email. You get a workspace right away.
Create accountIn the console, open API keys and create one. Copy it now: it is shown only once.
vb_live_…Add the MCP server to Claude Code, or call the REST endpoint from your server.
claude mcp add \
--transport http \
viberoo-search \
https://tools.viberoo.ai/api/v1/mcp \
--header "Authorization: Bearer $KEY"Still unsure? Create an account and ask the chat. The full API reference is in the console.
From a SearXNG instance that runs next to the API on a private network. SearXNG queries public search engines and merges their results.
It opens the top results (up to 5) in a sandboxed browser and returns each page's main text, up to 10,000 characters. It is slower: a few seconds per page, read in parallel. A page that can't be read returns content: null with a reason, and the search still succeeds.
Usage records keep identifiers and counts only: which key made the call, when, and how many results and pages came back. They do not keep the query text or page content.
Any MCP client that supports the Streamable HTTP transport, such as Claude Code and Cursor, and anything that can send an HTTPS POST with a bearer token.
Up to 10 results and 5 page reads per search, and 60 requests per minute per key by default. Over the limit you get 429 with a Retry-After header.
Treat it as untrusted third-party content. The web_search tool tells the model so, and the service itself never acts on page text, but your agent should not follow instructions it finds there.