Claude Code can write your refactors, squash your bugs, and scaffold an entire repo from a paragraph of instructions. Ask it what's on the Hacker News front page right now, though, and it stalls. It has no internet access. That's the wall every developer hits eventually: the agent is brilliant with code, blind to everything else. The fix is smaller than you'd think. NeuroAPI ships a hosted MCP server that exposes every web primitive — scrape, search, crawl, extract, interact — as typed tools an agent can call mid-conversation. You paste one URL into your config, and Claude Code suddenly has real-time access to the live web. No wrappers, no local browser pool, no glue code to maintain. I wired this up on a Saturday afternoon to help with a documentation migration project. By Monday I'd found uses for it I hadn't planned. Here's everything I learned. What you actually get NeuroAPI's MCP server exposes twelve tools, all prefixed with neuroapi_ to avoid collisions with other MCP servers you might have running. The ones that matter most in a coding context: neuroapi_scrape — fetch any URL as clean Markdown. This is the workhorse. One credit per call. neuroapi_search — web search that returns scraped content, not just links. Useful when you need Claude to research a library or API before writing code against it. neuroapi_extract — structured JSON from a page, driven by a schema you define. Two credits. This one's a sleeper hit for data pipelines. neuroapi_crawl — recursive site crawl with depth and path controls. Returns a job ID you poll with neuroapi_crawl_status. neuroapi_interact — stateful browser sessions. Scrape a page first, then click buttons, fill forms, and scroll in follow-up calls using the returned sessionId. The full list includes neuroapi_map, neuroapi_screenshot, neuroapi_summary, neuroapi_question, neuroapi_highlights, neuroapi_batch_scrape, neuroapi_parse, and neuroapi_balance. Each one is a single tool call from inside your agent session. Setting it up in Claude Code NeuroAPI uses Streamable HTTP transport, which means your API key goes in the URL path — no Authorization header needed. This matters because not every MCP client supports custom headers, but every one of them supports a URL. The config lives in your Claude Code MCP settings. Open it with: claude mcp add neuroapi --transport http https://neuroapi.me/api/nk_live_YOUR_KEY/mcp Replace nk_live_YOUR_KEY with your actual key from the NeuroAPI dashboard. If you prefer editing the JSON directly: { "mcpServers": { "neuroapi": { "url": "https://neuroapi.me/api/nk_live_YOUR_KEY/mcp" } } } That's it. Restart Claude Code and the tools appear automatically. You'll see them listed when Claude describes its available capabilities. No SDK installation, no npm install, no environment variables to wrangle. Real workflows that changed how I use Claude Code Research before writing code I was building an integration against a third-party API whose docs had moved three times. Instead of copy-pasting docs into the prompt, I told Claude Code: Go to https://api.example.com/docs, read the authentication section, and write me a TypeScript client for the /users and /orders endpoints. Claude called neuroapi_scrape, got the page content back as Markdown, parsed the auth flow, and wrote a working client. The code actually matched the current docs, not some cached version from a training run. When the API docs changed two weeks later, I re-ran the same prompt and got updated code. Crawling a site for migration work For a CMS migration, I needed to know every URL on the old site. I asked Claude to map it out: Use neuroapi_map to discover every URL on https://old-site.example.com. Group them by path prefix and count how many pages are in each section. Claude called the map tool, got back the full URL list, and organized it into a structured report. I then had it start a crawl on the top sections to extract front-matter data using neuroapi_extract with a JSON schema for title, description, author, and publish date. The whole pipeline ran inside the agent session. Debugging with live context This one surprised me. I was debugging a CSS layout issue and asked Claude to take a screenshot of the deployed page via neuroapi_screenshot, then describe what it saw. It identified the broken flex container immediately. Not a replacement for browser DevTools, but faster than explaining the layout in words. The interact workflow: stateful browsing Most MCP scraping tools are one-shot. You get the page, you're done. neuroapi_interact is different. When you call neuroapi_scrape, the response includes a sessionId. Pass that session ID to neuroapi_interact and you can click buttons, fill forms, scroll, and execute small scripts against the live page — all without leaving Claude Code. A concrete example. Say you need to check if a staging site's login flow works: You: Scrape https://staging.example.com/login, then fill in the email field with test@example.com and the password field with testpass123, click the Sign In button, and tell me what page loads. Claude will call neuroapi_scrape to load the page, then make a follow-up neuroapi_interact call with a prompt like "fill email with test@example.com, fill password with testpass123, click Sign In". The interact tool runs those actions in a real browser and returns the resulting DOM. Claude reads it and tells you what happened. Sessions expire after a short idle window, so don't try to keep one open for hours. If you get a session_expired error, just re-scrape the URL and continue. How it compares to self-hosted alternatives You could run your own MCP server with a headless browser, Playwright scripts, and a parsing pipeline. Some teams do. Here's the trade-off matrix: