Four steps, one tool at a time
- Find. Agents scan new GitHub projects and Show HN posts. For now they start with newer tools few people know yet, to show that such tools can be tested and trusted too. Anyone can also submit a public repo or a website. So far pickfit has read 1150 READMEs.
- Read. A model reads each README and writes a profile: what it does, where it runs, what it costs, whether it works offline. Every field quotes its source; a field with no source is left empty, not guessed. Claude does the reading now; tools listed before mid-November 2026 were read by Meta’s Muse Spark.
- Install. A verification agent installs the tool in a fresh sandbox (Linux in gVisor; some by hand on a real Mac), fixes what the README got wrong, checks the main features one by one, then replays the working steps in a brand-new sandbox. 729 tools passed a replay.
- Match. Your sentence and every verified profile that meets your conditions go to Claude in one call. Each reason it gives must name a profile field; code checks the value and drops any reason that doesn’t hold. If the model is unavailable, pickfit ranks by rules and says so.
What “verified” means
A tool card says how far the run got. Higher is stronger evidence:
- Installs and startsThe command runs and prints its version or help. A GUI app opened its own window.
- Its main function worksThe run gave it real input and checked the output: a file converted, a query answered.
- Its service answers requestsA local server came up and answered. For an MCP server, a client called one of its tools.
- It produces model outputA local AI tool loaded a model and generated text, speech or an image.
A run says where it happened (OS and CPU) and when. Its OS mark is green when the steps ran on the system you chose; otherwise the card says in words where they did run.
Web apps: no install, just open it
A web app runs on its makers’ site, so there is nothing to install. pickfit opens it once in a sandboxed browser: it counts as installed when the page loads to its own main screen (not an error page, a parked domain or a sign-in wall), and as working when a model uses it for what it claims and pickfit checks the result itself. It never posts, mails, signs up or orders on a real site. The browser’s own records say whether an account was needed, whether ads showed, how many other sites the page contacted, and whether what pickfit typed in left the device. Only an installable web app (a PWA) that pickfit reloaded with the network off counts as working offline. 100 web apps passed so far.
When a web app and a tool to install fit your need about equally, the web app comes first. A job that runs for long or in the background, ties into the system or works through many local files leans to the tool you install. If you ask for offline use or for your data to stay on your device, a web app qualifies only when its run showed exactly that.
What the labels mean
Stitching
The rufous running stitch on a tool card separates the claim from the run record. pickfit draws it only under evidence from its own run. A tool pickfit has not run, such as a well-known option Fitz mentions from general knowledge, never gets stitches, a badge, steps or a link.
Marks on a tool card
Facts every card repeats are shown as marks, each named on hover and for screen readers:
- macOS, Linux and WindowsWhere it ran
- Green OS marks: pickfit ran the install steps there. Grey ones: the README says it runs there, and pickfit has not tried.
- How far the run got
- Installs and starts, its main function works, its service answers, it produces model output. A window means a desktop app opened; a plug, an MCP client used it. The mark sits beside the run’s result, under “Why it fits and how we know” and in “How pickfit tested it”.
- Visual check
- pickfit looked at screenshots of the app. Green: it looks right. Amber with a number: things that look wrong; open it to see them. Grey and dashed: not checked yet.
- Known vulnerabilities
- Known high or critical vulnerabilities (OSV.dev) in what it installs. Green: none known. Red with a number: how many; open it for the list, and read it before you install. Grey and dashed: not checked, which does not mean safe.
- About the run
- The day it ran, and the version that ran.
- Source of a reason
- Every reason a pick fits rests on one profile field. Open the reason to see the field and the value it relies on.
Quality and popularity
Two scores out of 100, computed apart. Quality is a weighted mean: whether it installed (how many install actions, how many README fixes it needed) 20%, feature results (of the features with a verdict, the share that worked) 20%, how complete the install docs are 25% (a model grades them, quoting the README), maintenance 10%, issue responsiveness 10%, engineering hygiene 5%; reports from real use join as they accumulate. A part with no data drops out instead of counting as zero: a feature pickfit could not try does not lower the score, and only a failure caused by the tool itself does. What drops out lowers confidence, the share of the score that had data, shown beside it with what is missing. Popularity compares stars, forks, downloads and Hacker News points with other tools of the same kind. The example tapes above are illustrative.
A web app has no README to install from, so its quality is the browser run 50% (it loaded, or it worked when used) and what the browser saw 40% (no account needed, nothing you type sent off your device, no ads, works offline), with real use joining the same way. Its popularity is its GitHub stars, or the points of its Show HN post when it has no public code.
Desktop and web apps that pickfit opened also get a visual check: a model looks at pickfit’s own screenshots and lists defects you could point to, such as text cut off, something covering content, a clipped control, a word broken mid-word, low contrast, tiny text, raw code like >, a broken image, misalignment or crowding. Taste and style don’t count. If it finds a defect the app causes, the quality score drops by 5%; finding none adds nothing. An app not checked yet is shown as not visually checked and keeps its score. How serious each defect is doesn’t count yet: in a review by two independent checkers, about one listed defect in four was not real, and most defects called serious were minor.
pickfit also checks every package a fresh install brings against OSV.dev, the public database of known vulnerabilities. A known high or critical vulnerability in any of them caps quality at 79, so the tool can’t be a hidden gem: we would rather miss a gem than recommend software with a known hole. Software pickfit could not check (no lockfile to read, for example) keeps its score, and its page says it was not checked.
Hidden gem marks a verified tool with quality 80 or more and popularity 45 or less: 161 tools today. Fitz ranks by fit to your need, not by either score.
Features checked
The run tries each main feature the README advertises. Some need an account, a GPU, a display or hardware the sandbox doesn’t have; those stay unchecked, and the card gives the reason. A feature whose test failed with the tool’s own error is marked as failed.
Match labels
When the model is paused, pickfit ranks by rules and labels each pick Good match Partial match or Weak match by which of your words its profile mentions. With the model, Claude writes the fit sentence and code checks each of its reasons.
Fitz’s rules
pickfit is neutral on purpose. It belongs to no agent maker and no software vendor, takes no money for placement, and doesn’t make any of the software it suggests. That is what lets it say a little-known tool fits better, or that what you already have is enough.
Privacy: no account, no IP address stored. Your sentence goes to Anthropic’s Claude through the Anthropic API; pickfit keeps only the conditions it parsed, such as the task and the OS. Compare picks stay in your own browser.
- Tested before it’s suggested
- Ranked by fit, not stars
- Untested tools are labeled
- Says when built-in tools will do
- No ads, no referral links
Found a tool that should be here?
Authors and users can submit a GitHub project, or the address of a web app, right here on the site or through the MCP submit_software tool. It gets the same review: read, tried in a sandbox, scored. Submitting never earns a better rank.