Bright Data MCP
Search, scrape, extract, and automate the public web from MCP-compatible AI clients.
The evidence shows that enabled capabilities can be narrowed with GROUPS and TOOLS, while release workflows use read-only contents permission, OIDC, npm provenance, and signature auditing. The README also discloses that requests traverse Bright Data infrastructure and explains billing behavior. Deductions apply because five base tools are always enabled, browser actions such as clicking, typing, and submitting have no per-action confirmation mechanism, and there is no general undo facility. Tokens may be stored in environment variables but are also placed in URL query strings in several examples; the supplied files do not document log redaction, rotation, retention, or scraped-data handling. The registry workflow downloads and executes an unpinned latest publisher binary without a shown checksum. Author, repository, package, and MIT copyright attribution are broadly coherent, but the publisher registry identity is unverified, so attribution is not maximally strong.
The README, package manifest, and tests present a reasonably consistent product covering search, schema validation, and an stdio health path. Dependency versions, the Node launch command, and hosted/local deployment paths are documented. Tests demonstrate explicit handling of unknown datasets, invalid filters, non-JSON responses, and session_stats. Deductions apply because only a small fraction of the 69 tools is covered; the supplied evidence does not establish error behavior, retries, timeouts, or degradation for most tools or remote outages. Operation also depends on Bright Data services, an account, and external websites.
Audiences and scenarios are described thoroughly across research, commerce, competitor analysis, coding, social monitoring, and content creation. Configuration is supplied for many MCP clients and both hosted and local environments. Tool groups, URL validation, batch limits, filter depth, browser snapshot sequencing, and cheaper static-page alternatives provide useful boundaries and selection guidance. Deductions apply because base tools cannot be fully disabled, some boundaries are README assertions rather than corroborated implementation evidence, and login walls, site terms, regional constraints, and client-specific differences are not treated comprehensively.
The README has strong information architecture across quick start, pricing, scenarios, groups, and tool reference, with installation examples for numerous clients. The complete MIT text agrees with package metadata, justifying full license credit. Naming is generally stable, and the repository provides a version, issue path, and automated release route. Deductions apply because no changelog or migration history for the supplied revision is present, FAQ and troubleshooting coverage is limited, and known limitations are scattered rather than consolidated. Maintenance responsibility is inferred from the Bright Data author/copyright fields and issue URL, without named maintainers or a support commitment. The README's brightdata-com repository link also differs slightly from the stated brightdata object name.
Documented output types, field examples, batch limits, fallback guidance, and tool-selection rules make outputs readily usable in agent workflows. Combining search, scraping, structured extraction, and browser automation offers credible marginal value over a single-purpose fetcher. A free allowance, metered pricing, exhaustion stopping, and spend caps improve cost control. Deductions apply because many superiority and comparison claims are promotional and lack benchmarks or broad tests in the supplied files; per-GB browser pricing, per-record extraction, and large responses may impose financial and context costs.
Identity, version, license, dependencies, release mechanics, and a few behaviors can be cross-checked among the README, package manifest, workflows, and tests. Tests concretely support selected schema, parser-error, and health-interface claims. Deductions apply because CAPTCHA bypass, reliability across 69 tools, competitor comparisons, free-tier details, and broad platform output claims are not traced tool-by-tool to supplied implementation or tests. Promotional comparisons are not consistently separated from established facts, and static files cannot independently verify hosted-service behavior.
- Several quick-start examples place the API token in a URL query parameter. Query strings may be retained in client history, proxies, or server logs; confirm redaction and prefer a safer secret-delivery method before deployment.
- Browser tools can navigate, click, type, and submit, but the supplied material shows no step-level authorization or undo mechanism. Restrict enabled tools at the MCP client and require human confirmation for consequential actions.
- Every request traverses third-party hosted infrastructure. Independently verify retention, processing location, access controls, and compliance terms before sending regulated, personal, or confidential data.
- The registry workflow executes a binary downloaded from an unpinned latest URL without a shown checksum. Supply-chain review should pin the release and verify artifact integrity.
- Pricing, free allowance, platform coverage, and unblocking claims come from the README and are not independently established by the static evidence; verify them before adoption.
What does this agent do, and when should you use it?
Bright Data MCP is a public-web data server exposing 69 tools through the Model Context Protocol. Its interfaces cover Google, Bing, and Yandex search, page conversion to Markdown or HTML, AI-assisted JSON extraction, and interactive remote-browser sessions. Dedicated `web_data_*` tools return structured records from sources including Amazon, LinkedIn, Instagram, TikTok, YouTube, X, Reddit, Crunchbase, npm, and PyPI. Requests run through Bright Data's infrastructure for proxy rotation, CAPTCHA handling, bot-detection bypass, and geo-restricted access. Teams can connect to a hosted MCP endpoint or launch a local server with `npx @brightdata/mcp`; both modes require a Bright Data API token. It is a strong fit when broad, current web coverage outweighs the cost and operational dependency of using a managed data provider.
An MCP client calls search_engine or search_engine_batch for Google, Bing, or Yandex results, while discover finds and ranks sources by intent, date, and relevance. Known URLs go through scrape_as_markdown, scrape_batch, or scrape_as_html; extract turns an arbitrary page into structured JSON with an optional extraction prompt. Platform-specific web_data_* tools read supported Amazon, Walmart, LinkedIn, TikTok, YouTube, Crunchbase, npm, PyPI, and other resources and return defined data fields. For interactive pages, the server opens a remote session with scraping_browser_navigate, exposes element references through scraping_browser_snapshot, and then performs clicks, typing, scrolling, screenshots, text or HTML retrieval, and network-request inspection. Separate tools submit prompts to ChatGPT, Grok, and Perplexity and collect their responses as structured data. Results return to the calling client over MCP, while Bright Data handles the underlying public-web access and unblocking.
- A researcher working in Claude Code or another MCP client needs current search results, source pages, and relevance-ranked material rather than model-training knowledge.
- An e-commerce analyst needs price, availability, rating, review, seller, or product records from Amazon, Walmart, eBay, Best Buy, and Google Shopping.
- A competitive-intelligence team wants to combine funding, hiring, pricing-page, app-review, and social-media evidence into company profiles.
- A data engineer needs clean JSON from a supported platform or a custom structure extracted from an otherwise unsupported webpage.
- An automation developer must collect content from dynamic pages that require navigation, clicks, typing, scrolling, or screenshots.
- A brand or GEO team wants to compare how ChatGPT, Grok, and Perplexity describe a company, cite sources, and make recommendations.
What are this agent's strengths and limitations?
- One MCP service exposes 69 tools, including search, batch scraping, 45 platform-specific structured extractors, and 13 remote-browser operations.
- Proxy rotation, CAPTCHA handling, bot-detection bypass, and geo-restriction handling are built into requests, avoiding separate proxy and headless-browser maintenance.
- It supports both a no-install hosted endpoint and a local
npx @brightdata/mcpserver, with tool groups for controlling the client context. discoveradds intent-based, date-filtered relevance ranking, while batch tools process up to 10 searches or pages per call.- The recurring free tier includes 5,000 requests per month without a credit card and includes browser automation.
- Web access depends on Bright Data and a valid API token; running the MCP server locally does not create an independent or offline scraping stack.
- Requests stop when the free allowance is exhausted unless the account has deposited funds, after which pay-as-you-go charges can apply.
- Specialized
web_data_*tools enforce URL shapes—for example, Amazon product URLs need/dp/and Walmart product URLs need/ip/. - Loading the complete catalog increases the tool context an AI client must process, so adopters may need to manage
GROUPSorTOOLScarefully. - Complex sites can require longer client timeouts, and local deployment adds a Node.js runtime requirement.
How do you install or deploy this agent?
The hosted option requires no server installation. Obtain a token from Bright Data account settings and register https://mcp.brightdata.com/mcp?token=YOUR_API_TOKEN_HERE as an MCP server. For Claude Code, run claude mcp add --transport http brightdata "https://mcp.brightdata.com/mcp?token=YOUR_API_TOKEN". For a local deployment, install Node.js and add {"mcpServers":{"Bright Data":{"command":"npx","args":["@brightdata/mcp"],"env":{"API_TOKEN":"<your-api-token-here>"}}}} to the client's MCP configuration. Limit the exposed surface with groups and tools URL parameters on the hosted service, or the GROUPS and TOOLS environment variables locally.
How do you use this agent?
After connecting, use search_engine to find information and scrape_as_markdown when the target URL is already known. Use search_engine_batch or scrape_batch for up to 10 queries or URLs per call. Choose discover for intent-aware research with relevance scoring and date filters, and prefer the matching web_data_* tool when the source platform is supported and structured JSON is desired. Enable the advanced_scraping group for extract and scrape_as_html. For interactive pages, call scraping_browser_navigate, then scraping_browser_snapshot, and act on the latest references with commands such as scraping_browser_click_ref or scraping_browser_type_ref; take another snapshot whenever the page changes. If a specialized web_data_* request fails, retry the same URL with scrape_as_markdown.
How does this agent compare with similar options?
Compared with the README's description of typical web MCP servers offering 2–10 mostly generic scraping tools, Bright Data MCP provides 69 tools, including 45 platform-specific JSON extractors, 13 remote-browser operations, three search engines, batching, and geo-targeting. Its stated differentiator is Bright Data unblocking on every request, plus capabilities such as LLM-response collection and npm/PyPI package data. The tradeoff is dependency on Bright Data accounts, tokens, service availability, and usage pricing.
FAQ
What does the free tier include, and what happens afterward?
Do I have to host the MCP server myself?
npx @brightdata/mcp. Both options use Bright Data's underlying web-access service and require an API token.Why did a web_data_* tool return no data?
/dp/ for an Amazon product. If the specialized call still fails, use scrape_as_markdown on the same URL.Must every client load all 69 tools?
discover tools are always available. Other capabilities can be enabled by group, such as ecommerce, social, browser, or code, or selected individually with TOOLS.Can it handle pages that require interaction?
browser group supports remote navigation, snapshots, clicks, typing, scrolling, waiting, screenshots, and network-request inspection. Reference-based actions must use the latest snapshot, so a changed page needs a fresh snapshot.