AI and agentsProvider guide

Bright Data MCP Server for AI Agents: Tools and Setup

Bright Data's MCP server from its docs: 32 tools in groups, the 5,000-request free tier, remote and local setup, and settings that limit cost and context.

Checked against provider docs 5 min read
How this page was made. Based on each provider's public documentation, checked on . We haven't run these through our hands-on tests yet. Our verdicts on each provider's core products are in the reviews.

Short answer

Search, scrape, pre-built scrapers and a remote browser in one server

Bright Data's MCP server runs on its Web Unlocker and exposes 32 tools, from page scraping and multi-engine search to pre-built scrapers and browser control. It includes 5,000 requests a month free, and you can load only the tool groups you need.

Bright Data's MCP server gives AI agents a way to search the web, scrape pages and run Bright Data's pre-built scrapers from inside Claude, Cursor or a custom agent. This page covers what the server offers, how to connect it and which settings matter, using only Bright Data's own documentation.

What it is

Bright Data's docs describe the server as giving "AI agents and assistants real-time access to public web data" through MCP (overview). It runs on the Web Unlocker API, which the same page says handles proxy rotation, anti-bot challenges and CAPTCHA solving automatically. The code is on GitHub under the MIT license.

If MCP itself is new to you, start with what an MCP server is.

Free tier and billing

The free-tier page sets out how it works:

  • 5,000 requests a month. The pool is shared with Web Unlocker API, SERP API, Web Scraper API, Scraper Studio and Browser API.
  • Monthly reset. Credits reset to 5,000 on the first of each month, and unused ones don't roll over.
  • After the allowance. With funds on the account, usage moves to your plan's pay-as-you-go rates. Without funds, requests return an error.
  • Rate limit on unfunded accounts. 1,000 requests per minute across SERP API, Web Unlocker API and proxy products, removed once you add funds.

The free-tier page doesn't list a per-request price after the allowance, so check Bright Data's pricing pages for that. The tools page notes that run_scraper is billed per record.

The tools

The tools page lists 32 tools, all enabled by default, in these groups:

GroupWhat it addsExample tools
Always onSearch, single-page scraping and AI-ranked discoverysearch_engine, scrape_as_markdown, discover
advanced_scrapingBatches, raw HTML and AI extraction to JSONsearch_engine_batch, scrape_batch, scrape_as_html, extract, session_stats
scrapersPre-built scrapers for named sitessearch_scrapers, get_scraper_details, run_scraper, get_scraper_results
social, businessSearchable datasetslist_dataset_fields, search_dataset
browserA remote browser the agent can drivescraping_browser_navigate, scraping_browser_click_ref, scraping_browser_type_ref, scraping_browser_screenshot
geoAnswers from other AI assistantschatgpt_ai_insights, perplexity_ai_insights

A few details from the same page:

  • search_engine covers Google, Bing and Yandex.
  • search_engine_batch and scrape_batch take up to 10 queries or pages at once.
  • The browser tools work from an ARIA snapshot: scraping_browser_snapshot lists interactive elements with refs, and the click and type tools act on those refs.
  • The FAQ names Amazon, Booking.com, LinkedIn and Zillow among the platforms with pre-built scrapers.

Connect it remotely

The remote server needs no install. The remote quickstart gives two endpoints:

Streamable HTTP:  https://mcp.brightdata.com/mcp?token=YOUR_API_TOKEN
SSE:              https://mcp.brightdata.com/sse?token=YOUR_API_TOKEN

The quickstart says no username is needed, only the API token, and that new users get the token in their welcome email. In Claude Code, the README gives this command:

claude mcp add --transport http brightdata "https://mcp.brightdata.com/mcp?token=YOUR_API_TOKEN"

Connect it locally

For a local server in Claude Desktop, add this to claude_desktop_config.json (GitHub):

{
  "mcpServers": {
    "Bright Data": {
      "command": "npx",
      "args": ["@brightdata/mcp"],
      "env": {
        "API_TOKEN": "<your API token>"
      }
    }
  }
}

Settings worth knowing

The README documents these environment variables:

  • GROUPS and TOOLS. Load only some groups or named tools, for example GROUPS=scrapers,browser. The overview says this reduces context size. On the remote server, pass groups and tools as URL parameters.
  • RATE_LIMIT. Cap usage, for example 100/1h. The default is unlimited, so set this if an agent will run unattended.
  • WEB_UNLOCKER_ZONE and BROWSER_ZONE. Use your own zones instead of the defaults, mcp_unlocker and mcp_browser.
  • CONTENT_GATE. On by default. The server asks you before returning page content that looks like instructions to change the MCP configuration, with a 120-second default timeout.
  • BROWSER_FILTERED_SNAPSHOT. Limits browser snapshots to buttons, links and fields. Off by default.

Where it fits

Bright Data's server puts multi-engine search, pre-built scrapers for named sites and a remote browser in one place, per its tools page. The comparison of MCP servers shows it next to Oxylabs and ZenRows. For our verdict on Bright Data's core proxy and scraping products, see the Bright Data review.

FAQ

Is the Bright Data MCP server free?

The server code is open source under the MIT license, and Bright Data's docs include 5,000 requests a month free. That allowance is shared with Web Unlocker, SERP API, Web Scraper API, Scraper Studio and Browser API, resets on the first of each month, and does not roll over (free tier).

What happens when the free requests run out?

Per the free-tier page, if you have deposited funds, usage moves to your pre-paid balance at your plan's pay-as-you-go rates. With no funds, requests return an error until you add some.

Can I limit which tools the agent sees?

Yes. Set GROUPS to a comma-separated list such as scrapers,browser, or TOOLS to name individual tools. The remote URL accepts the same as groups and tools parameters. Fewer tools means less of the model's context is used.

Which AI apps does it work with?

The GitHub README lists Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Gemini CLI, Zed and Warp, and says any client that supports local MCP servers works through npx @brightdata/mcp. The FAQ also mentions LangChain and custom agents.

Advertisement

Sources