pickfit

pickfit is a neutral software advisor for people and their AI agents. It never recommends a tool it hasn't installed and tried. The problem: AI coding agents made building software cheap, so new tools ship every day on GitHub, Show HN, Flathub and the web. Finding the one that fits got harder. Search and directories reward popularity, models don't know what shipped after their training cutoff, and small projects often have install steps that don't work. So people and agents rebuild what already exists, and pay for it later in maintenance and tokens. What pickfit does: it discovers tools (dev CLIs, local AI tools, MCP servers, Linux and macOS desktop apps, web apps), reads each README with quote-checked extraction, installs it in a fresh sandbox, has a Meta model operate it to check the claimed features really work, then replays the steps in a new box with the network off. Only then can it be recommended. Results are ranked by fit (OS, CPU, free, offline, data stays local), not stars, with no ads or referral links; lesser-known tools that do the job well are surfaced as hidden gems. Use it at https://pickfit.app (English and Traditional Chinese) or from any agent through the MCP server at https://pickfit.app/mcp: find_software, find_alternatives, get_install_guide, report_outcome, submit_software. Makers can also paste a link at https://pickfit.app/submit and watch it go through the same tests. Every answer carries how_we_know and an evidence_url. As of 5 Nov: 1,050 projects read, 732 verified at least through install, 647 replayed in a fresh sandbox, 85 web apps tried in a browser, 159 hidden gems. Disclosure: the idea and planning notes predate the event and are not in the repository; all code, data, the website, the mascot and the video were made during the hackathon. Full report (hypotheses, every experiment including failures): https://ncchen99.github.io/pickfit/