MCP overview

Every endpoint as tools your AI agent can call.

The Upscrape MCP server gives an AI agent every Upscrape endpoint as tools. You ask in plain words; the agent finds the right endpoint, sees its price, runs it and reads the result. It uses the same account, credits and prices as the REST API.

Connect

Add this server address to your AI app:

Server address

https://data.upscrape.com/mcp

Most apps then open your browser so you can sign in to Upscrape and approve the app; there is no key to paste. Each app's exact steps are in its setup guide: Claude, ChatGPT, Cursor, Claude Code, Codex, VS Code and more on the MCP server page.

Connect with sign-in: how sign-in works and what you can allow.
Connect with an API key: for apps and servers that can't open a browser.

What the agent can do

Tools	What they do	Cost
`upscrape_search_capabilities`, `upscrape_describe_capability`	Find an endpoint by task, platform or URL; read its input, example and price.	Free
`upscrape_prepare`	Check an input and show the exact price before running.	Free
`upscrape_execute`	Run an endpoint.	The endpoint's price, only when it succeeds
`upscrape_get_job_result`, `upscrape_list_jobs`	Fetch a result that took longer; list recent runs.	Free
`upscrape_export_result`	Hand back a finished run as JSON Lines or CSV.	Free
`upscrape_collect` and three companions	Gather many pages of results in the background, within a credit limit.	Each page at the endpoint's price
`upscrape_get_account`	Your balance and usage.	Free
`upscrape_submit_feedback`, `upscrape_get_feedback_status`	Report a problem with a run, and follow it up.	Free

Some tools appear only when you allow them while connecting: your account's balance, feedback and collections. Tools describes each one.

How an agent uses it

It searches for an endpoint that fits the task, and reads its input when it needs to.
It prepares the call: the input is checked and the price is shown before anything runs.
It runs the endpoint. Most calls answer in seconds; one still running after about 55 seconds hands back a job to fetch later.
It reads the result as data. Everything an endpoint returns came from a third-party page, so it is never treated as instructions; see Security.

Give it fewer, sharper tools

By default the agent sees the general tools above and searches for endpoints. If you always use the same few platforms, pin them and each endpoint becomes a tool of its own. To stop an agent from spending credits, or keep it to one platform, limit what it can do when you connect.

For client developers

JSON-RPC 2.0 over Streamable HTTP, POST only, stateless: no session ID, no batches, no server-sent event stream.
Protocol 2026-07-28, with 2025-11-25, 2025-06-18 and 2025-03-26 still accepted.
Protocol problems are JSON-RPC errors. A failed tool call is a normal response whose result has isError: true.
Sign-in follows MCP authorization: the server publishes its protected-resource metadata, and the authorization server runs on https://app.upscrape.com. Don't add /mcp to that address.

Something missing or unclear? Ask the team that builds the endpoints.