7 MCP servers that give Claude live web data

Claude's training data has a last day, and every question about this week falls behind it. MCP servers (small add-ons that give Claude new tools) fix that. The trick is knowing which job you have: fetch, search, scrape, docs or browser.

How we picked: each repo had a commit in the last 90 days (checked November 7, 2026) and installs from its own README. Any paid API behind it has a free tier confirmed on vendor pricing this month. We dropped jina-ai/reader as stale.

The 7 MCP servers at a glance #

1. Fetch: the zero-setup default #

The official Fetch server from the MCP team lives in the reference monorepo (91,846 stars).

What: one tool that fetches a URL and returns its content as markdown — no account, no API key. Why: the fastest way to let Claude read today's page. Use when: you already have the URL and just need the content. Install:

uvx mcp-server-fetch

Catch: one URL at a time — no search, and JavaScript-heavy pages can come back thin.

2. Tavily: search built for LLMs #

Tavily exposes its search API as a remote MCP server, so there is nothing to run locally (free tier).

What: 1,000 free credits a month (1 credit per basic search), no card, over a hosted endpoint. Why: results arrive cleaned and ranked for model consumption, not as raw SERP HTML. Use when: your agent needs current web results with the least plumbing. Install:

claude mcp add --transport http tavily "https://mcp.tavily.com/mcp/?tavilyApiKey=<your-api-key>"

Catch: advanced searches cost 2 credits, and when the 1,000 run out you wait for the monthly reset.

3. Exa: find pages by meaning #

Exa is a meaning-based search index behind a hosted MCP server. It works without an account at low rate limits; sign in for more (free tier).

What: a free balance that resets to $10 on the 1st of each month, no card — search, fetch and crawling included. Why: semantic queries surface the page you mean, not just pages containing the keywords. Use when: you research questions where keyword search returns noise. Install:

claude plugin install exa@claude-plugins-official

Catch: at $4–7 per 1,000 searches the $10 buys roughly 1,400–2,500 queries, not unlimited.

4. Brave Search: an independent index #

Brave's official server wraps the Brave Search API (free tier).

What: $5 of free credits each month — about 1,000 searches at $5 per 1,000 requests. Why: an index independent of Google and Bing, useful as a second opinion. Use when: you want to cross-check another search tool. Install:

npx -y @brave/brave-search-mcp-server   # needs BRAVE_API_KEY

Catch: a credit card is required at signup — Brave says it is for abuse prevention.

5. Firecrawl: whole sites to markdown #

Firecrawl offers scrape, crawl, map and search endpoints that return clean markdown (free tier).

What: 1,000 free credits a month (1 credit per page), no card. Why: turns many pages from one domain into model-ready text in one call. Use when: Claude needs more than one page from the same site. Install:

env FIRECRAWL_API_KEY=fc-YOUR_API_KEY npx -y firecrawl-mcp

Catch: 2 concurrent requests, credits do not roll over, and error pages still cost one.

6. Context7: current docs, not stale weights #

Context7 (64,309 stars) serves up-to-date library documentation to any MCP client.

What: pulls current docs for your libraries into the prompt. Free; a free API key raises the rate limits. Why: Claude stops writing code for an old version of the API. Use when: Claude writes code against a library whose API has moved. Install:

claude mcp add --scope user context7 -- npx -y @upstash/context7-mcp --api-key YOUR_API_KEY

Catch: coverage is community-submitted, so niche libraries may be missing.

7. Playwright: when pages need a browser #

Microsoft's official server (Apache-2.0, 39,289 stars) lets Claude drive a real browser.

What: click, type, scroll, screenshot and read the DOM through structured accessibility snapshots. Why: reaches logins, infinite scroll and JS apps that fetch and search cannot. Use when: the data only exists after a browser renders it. Install:

claude mcp add playwright npx @playwright/mcp@latest

Catch: it runs a local browser, so it is slower and more brittle than an HTTP call.

Which one to try first #

Just want Claude to read a URL → #1. Need current answers → #2. Research where keyword search fails → #3. A second search opinion → #4. Many pages from one site → #5. Coding against fast-moving libraries → #6. Data behind a login or JS app → #7.