Skip to main content
Ask your agent a question in plain language and get back a table built from live web search results. The agent writes one read-only SELECT in the DuckDB dialect, following the guide that ships with the tool. The server runs the searches, fetches the pages, extracts the fields the query asks for with an LLM, removes duplicates and returns a table. One tool call covers many pages and returns structured rows. Your agent doesn’t have to read results one at a time and assemble them by hand. SELECT is a separate MCP server from the Keenable MCP server, with its own URL. Adding it doesn’t change the search_web_pages and fetch_page_content server, and you can run both side by side:
The server also answers on https://webql.keenable.ai/mcp, its first address. Neither address redirects to the other, because a redirect ends an MCP session. A client set up on the old address keeps working. Set up new clients with select.keenable.ai. Any Keenable account can connect. It’s the same account you use for the console and API keys. If you don’t have one, sign up at app.keenable.ai/signup. The account also gives you the SELECT playground at app.keenable.ai/select/start.

Connect

You can connect with a browser sign-in or with an API key in a header. Which one you can use depends on the client:

Claude Code

Browser sign-in, or your own key in a header

Claude

Desktop, claude.ai and mobile, from the connector directory

ChatGPT

From the Plugins Directory, or by URL in developer mode

Codex, Cursor

Any client that takes a URL and a header

One click

Cursor and VS Code add the server from a link, without a key. The client asks you to sign in to Keenable the first time you use the server.

Claude Code

Sign in through the browser, with no key to paste:
Then run /mcp in a session, pick keenable-select, and choose Authenticate. Your browser opens a Keenable sign-in; approve it and the tools appear. To use an API key instead, add the header and skip the sign-in. Create the key in the console:
--scope user makes the server available in every project. Without it the entry is written for the current directory only.

Claude Desktop, claude.ai and mobile

Keenable SELECT is listed in Claude’s connector directory, next to the Keenable search connector. The connector signs you in and doesn’t take a key. If you need a specific API key applied, use Claude Code or another client that sends headers.
1

Add the connector

Open Customize → Connectors, click +, and search for Keenable SELECT. There is nothing to paste.If it isn’t listed in your account yet, add it by URL instead: Customize → Connectors → + Add → Add custom connector. Under Authentication choose Sign in now, and under OAuth client choose Register automatically. Name it Keenable SELECT so you can tell it apart from the Keenable search connector, which is a separate server with different tools. Use this URL:
2

Connect

Click Connect. Your browser opens a Keenable sign-in; approve it and the tools appear.
3

Enable it in a chat

Click + in the lower left of the message box, hover Connectors, and toggle it on for the conversation.

ChatGPT

Keenable SELECT is a published plugin in OpenAI’s Plugins Directory, the catalog ChatGPT shares with Codex. It is a second listing next to the Keenable search plugin, because the two are different servers with different tools.
1

Install the plugin

Open the listing, or find Keenable SELECT in the directory, and install it. There is no config file, no developer mode, and no URL to paste.
2

Sign in

The plugin signs in to your Keenable account through OAuth. Click through the sign-in prompt the first time you use it. Usage is metered against the organization of the account you connect.
3

Use it in a chat

Pick the plugin from the composer’s tools menu before you ask your question. A query can take seconds to minutes; see What to expect.
If the plugin isn’t in the directory for your account yet, add the server by URL.
  1. Settings → Connectors → Advanced settings → Developer mode. Without it, ChatGPT offers no way to add a server by URL.
  2. Settings → Connectors → Create, paste https://select.keenable.ai/mcp, and choose OAuth as the authentication. Sign in when the browser opens; there is no client ID or secret to fill in.
Deep research in ChatGPT accepts only connectors that expose search and fetch tools. This server exposes select instead, so use it as an ordinary plugin, and the Keenable search plugin for deep research.

Codex, Cursor and other clients

Codex can also install Keenable SELECT from the Plugins Directory it shares with ChatGPT, with an OAuth sign-in. To use your own API key instead, add the server to config.toml as shown below. Any client that takes a remote MCP URL and a header works. Create a key in the console first. For Codex, add this to ~/.codex/config.toml:
For Cursor, Devin Desktop (formerly Windsurf) and similar clients, use this shape:
If your client offers only an Authorization header field, send Bearer keen_<your_key> in it instead.

Over HTTP, without an account

An agent with no Keenable account can run a SELECT query over plain HTTP and pay for it per query: $0.10 in USDC on Base with x402 at POST https://api.keenable.ai/v1/x402/select, or from a prepaid balance through ZeroClick. If a query fails, you aren’t charged. The request body, response fields and payment flow are on Pay per request. With an account, use the MCP server above: SELECT has no REST endpoint that takes an API key.

The tools

The main tool is select. Your agent fills in its parameters. They’re listed here so you can read a transcript and see what ran.
string
required
A single read-only SELECT in DuckDB dialect. Web pages enter the query through the WEB_SEARCH(...) and WEB_FETCH(...) table functions; the SEM_* functions extract, filter, rank and normalise values with an LLM.
boolean
Return a small sample of rows instead of the full table. Defaults to false.
You don’t need to learn this dialect or teach it to your agent. The select tool’s description is its full usage guide, with the syntax, the semantic functions and worked examples. Your agent reads it as soon as you enable the connector. To debug a query or steer your agent, see the function reference. Two more tools publish what a query found as a web page: generate_html_report builds a document from a brief plus the result sets you name and returns its link, and revoke_html_report takes that link back down. When a select call returns rows, the result can end with one extra line of text after the table. It names the result set and tells the agent it can call generate_html_report to get a link to share. The link is unlisted by default, and revoke_html_report turns it off. The line never changes the table or the result set. Organizations on zero data retention never get it. You ask “which electric cars sold in Europe come up most often, and what are their range and starting price?”, and the agent turns that into a query like this:
That is one tool call. It ran seven searches, merged 1,400 results into 1,125 unique pages, extracted three fields from each and grouped them, and came back with this (first 8 of 15 rows): This is a real run. The figures are whatever the pages said: the server reports its sources but doesn’t check them. See What to expect. The guide tells the agent to search with different phrasings, not variations of one. The server merges the returned pages and removes duplicate URLs before the rest of the query runs. The rows arrive in a compact tabular encoding instead of JSON, so a table this size takes up little of your agent’s context.

Result sets

Every call stores its output and returns a result set id, such as ra84968f0532 for the query above. A later query reads the stored rows when it puts that id in FROM, so a follow-up question needs no new searches. Ask “now only the ones under €30,000, longest range first” and the agent queries the stored table instead of going back to the web:
This query returned in seconds, because it searched and fetched no pages. The server also stores the searched pages as their own result set. Your agent can ask a different question of the same 1,125 pages without paying to search and fetch them again. Result sets are kept for 30 days.

Function reference

Use this table to read a query your agent wrote, or to tell the agent which function to use. Two table functions bring pages into a query, and five semantic functions work on each row: published_after and published_before take an offset such as 2h, 1d, 2w or 1mo, or a date such as 2026-07-01. A date filter drops pages that have no publish date. On SEM_EXTRACT and SEM_EXTRACT_ALL, evidence := true adds the verbatim quotes that each value came from. The rest of the syntax is DuckDB SQL. Every query reads FROM a WEB_SEARCH(...) call, a WEB_FETCH(...) call or a result set id. A query can’t return the full content column: extract from it, or take a bounded part such as substr(content, 1, 10000). The full guide your agent reads is at select.keenable.ai/guide.md.

Recipes

This query counts how many pages mention each restaurant:
You get one row per restaurant with an address and the number of pages that name it, most-mentioned first. This query enriches a stored result set. WEB_SEARCH takes a column as its argument, so it runs one search for each jacket in the result set ra1b2c3d4e5f:
You get one row per page that states a price, with the quote the price came from in price_evidence and the page’s URL. To add the same fields to every item of a list you already have, see Enrich a list with SELECT.

What to expect

  • A query takes seconds to minutes, because it runs real searches, fetches pages and calls an LLM for each row. A long wait doesn’t mean the call is stuck.
  • If a table is larger than about 50,000 characters, the response contains only a few sample rows, with long strings clipped. The full table stays available through the result set id.
  • A report takes minutes to build, and a client that gives up on a slow tool call cancels it. To make the report finish before the client stops waiting, build it from aggregated result sets, not from a raw table of hundreds of rows.
  • Extraction doesn’t fact-check its sources. A single page can be outdated or describe the wrong edition, so cross-check precise numbers across independent pages.

Errors

  • 401: no credential reached the server, or the credential is invalid. Check the header name and the key.
  • 403: this deployment runs an allowlist and your account isn’t on it. Production doesn’t run an allowlist, so you shouldn’t see this. If you do, email support@keenable.ai with the address you signed in with. Signing in again doesn’t clear it.
  • 429: too many queries are running at once, across the cluster or on your account. Retry shortly.
  • 503: the server couldn’t check your credential. Retry shortly.

Production checklist

Ship Keenable search and fetch to production: keys from the environment, rate limits, retries on 429 with Retry-After, stopping on 402, timeouts and status.

Time Machine

Search and read the web as Keenable captured it at a past moment with query_time. Nothing published later can appear. Early Access.

Monitors

Get new pages about any topic as Keenable crawls them: over MCP, by webhook, or in scheduled runs. Free for now. Quickstart, webhooks, schedules and limits.

Pricing

Search and fetch are free without a key. With a key, you get 100,000 free requests a month, then pay $4 per 1,000. What each call costs and what isn’t billed.