ToolRouter

"the whole site as markdown eats the window fast." That's a reply this week to a post about wiring Firecrawl into Claude Code, and it's the complaint I hear most about giving agents web access generally: the tool works, then the context window is gone after two pages. Firecrawl just raised 75 million in a Series B led by Smash Ventures, with YC, Altos Ventures, Nexus Venture Partners, Freestyle and Offline Ventures all participating, on top of an earlier 14.5 million Series A. The pitch is web data on tap for agents, and the product genuinely works: multi phase page loads, dynamic elements, PDFs, all handled. But the pricing is credits, not value. A basic scrape is 1 credit a page whether the page is 200 words or 20,000. JSON extraction adds 4 credits a page. A request that comes back as a 4xx or 5xx still gets billed. Crawl a whole site and every page lands in the model's context as markdown, whether anything on it mattered or not. That's the exact problem DeepScrape is built around. It generates a deterministic CSS selector once per page type, not once per request, and that selector self heals when the page changes instead of asking an LLM to re-read the whole document on every single pass. The model only gets pulled in for the narrow part: generating the selector, validating it, adapting it when a layout update breaks it. Everything else runs as plain extraction, not token spend. There's an MCP server so Claude and Cursor can call the scrapers directly as tools, and a site to MCP mode that turns any website into a structured data source an agent can query instead of re-scraping it cold every time. It's open source. github.com/stretchcloud/deep… teapaw.org/dng_notes/status/21154…