Open Source Radar — November 21: Firecrawl runs on your laptop

Today's board has one theme: the work moves onto your own machine. A scraper that no longer needs a server, a compliance check that runs before the upload leaves your laptop, and the inference toolkit for the NPU you already own.
Firecrawl (TypeScript, 205,600 stars, AGPL-3.0) shipped a local binary this month that does the scraping on your own hardware. Published to crates.io on November 11 and already at 3,687 downloads, it scrapes, crawls and maps sites without a database, queue or container, renders JavaScript in the Chrome you already have, parses PDFs and Office files locally, and serves the same v2 API on localhost so code written against Firecrawl runs unchanged and local routes need no key. One command looks for Claude Code, Claude Desktop, Cursor, Windsurf, Gemini CLI and Codex on the machine, backs up each config file, and adds the binary to all of them as an MCP server; search and the 550-plus data providers in Firecrawl's Alexandria stay cloud-only. It picked up 1,241 stars today after the README was reordered to lead with the local build, which tells you what the audience for a scraping API actually wanted: the crawl on the laptop, the account optional. Worth reading the "what runs where" table before you assume nothing leaves the machine — the split is honest but real.
app-store-compliance (Shell and Python, 2,900 stars) is a pre-submission guard for the week you lose to a one-line rejection. It documents 117 rejection patterns, each mapped to the guideline and the fix, plus the privacy manifest and Google Data safety checks that trip up apps built fast, and 95 tracked regulatory deadlines from the EU AI Act to COPPA. Installed into any of 13 coding agents — Claude Code, Codex, Cursor, Copilot, Gemini CLI, Windsurf and others — it hooks the submit command and stops the upload when a critical risk is still standing, with a read-only audit path for people who don't write code. The repo's urgency case is that AI-assisted building has flooded the stores while the rules tightened: it cites Apple figures of 7.77 million submissions reviewed and 1.93 million rejected, roughly one in four, and Google blocking 1.75 million Play submissions in 2025, with TechCrunch reporting 2026 release volume up about 60 percent year over year. Two caveats: the licence is a custom OpenRoots Agent License rather than a standard open-source one, and 869 stars in a day means most of those installs are untested — run the audit before you trust the hook.
OpenVINO (C++, 11,800 stars, Apache 2.0) is Intel's inference toolkit, and it is on the daily board for the least glamorous reason: it is the shortest path from a PyTorch, ONNX or JAX model to something that runs on the CPU, integrated GPU or NPU in a laptop you already own. The September 2026.4 release brought Gemma 4 12B, Qwen3-VL-4B with EAGLE-3 speculative decoding, Muse Glimmer 30B and DeepSeek OCR-2 to CPU and GPU, and FLUX.2-Klein 4B to the NPU, with backends for vLLM, ONNX Runtime, torch.compile and Keras 3 so you rarely rewrite the app. Its 217 stars today are not a launch spike — this is an eight-year-old project — but NPUs shipping as standard in consumer laptops made it the answer again to "can this run locally". Boring beats clever when the goal is inference on hardware you don't rent.
Worth watching this week: whether Firecrawl's local-first split survives AGPL scrutiny, and how many of those 13-agent installs hold up against a real submission.
Three tools that all shorten the distance between "works on my machine" and "shipped" — which one is still missing from your pipeline? Write to us via our contact page.







