Skip to main content

Web search

Web search. In your CLI, over MCP, your way.

Search the live web and get back text, not a page of links. Connected once, it answers your terminal and every assistant you already use, under one permission.

Results come back as title, url, date and the readable text of the page, which is what an agent needs and what a person reading a terminal wants too. Searches cost money, so they are the clearest thing on the daily ceiling.

A real command line for web search

Not a chat window pretending. Every argument is a flag, output is text a pipe can read, and each call lands on your record with the daily ceiling watching.

Connect the account once, in the browser, and the terminal has it from then on. No key on your laptop, nothing in your shell history worth stealing.

Web search from the command line

Or hand it to the assistant you already use

Claude Code, Cursor and Codex all speak MCP. Point any of them at one endpoint and they get the same web search tools, on the same permission and the same record as your terminal.

Web search over MCP, tool by tool

claude mcp add gbrain https://gbrain.io/mcp --header "Authorization: Bearer <token>"

Your assistant gets exa_search. The token comes from the client's own page in the workspace, minted once and shown once.

Connect it once, use it everywhere

Their consent screen, one click, and one connection answers all of these. No key is pasted anywhere, and nothing runs before you allow it.