SkillAgentSearch skills...

pickfit

Every tool, tested first. A software advisor for people and AI agents: pickfit installs and tests software in a sandbox before recommending it, over MCP and at pickfit.app. Built with Meta's Muse Spark models.

Install / Use

claude mcp add ncchen99 -- npx -y github:ncchen99/pickfit

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

77/100

Category

Automation

Supported Platforms

Claude Code
Claude Desktop

Our assessment of pickfit

pickfit scores 77/100 on our quality scale, 2503rd of 2,771 Automation skills we index.

Its MCP Server is 8.4 KB long, split into 7 sections with 3 code examples: a thorough specification that gives an agent plenty to work with.

It has 3 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
29/30
Structure
16/20
Description
15/15
Adoption
3/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated today, so pickfit is actively maintained.
  • It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 92/100, with 1 caution from licensing, adoption, age or documentation. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

pickfit compared with similar skills

All 4 of these similar skills score higher than pickfit; compare them before choosing.

SkillScoreStarsUpdatedFormat
pickfit (this skill)by ncchen99773todayMCP Server
Agent-Reachby Panniantong100113.3k12d agoCLAUDE.md
hyperframesby heygen-com10087.1ktodayCLAUDE.md
rufloby ruvnet10082.1ktodayCLAUDE.md
headroomby headroomlabs-ai10080.2ktodayCLAUDE.md

Frequently asked questions

How do I install pickfit?
Run claude mcp add ncchen99 -- npx -y github:ncchen99/pickfit. The install tabs above show the steps for each supported agent.
Which AI agents does pickfit work with?
It is written for Claude Code and Claude Desktop, as a MCP Server file. Other agents that read the same format can often use it too.
Is pickfit safe to use?
It is MIT-licensed and scores 92/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is pickfit still maintained?
The repository was last updated today, so pickfit is actively maintained.
<p align="center"> <a href="https://pickfit.app"><img src="worker/public/assets/og/pickfit-en.png" alt="pickfit: every tool, tested first. Fitz the tailorbird stands on a result tag: installed in a sandbox, features checked, replayed offline." width="860"></a> </p> <h1 align="center">pickfit</h1> <p align="center"><b>Every tool, tested first.</b><br> A software advisor for people and their AI agents. pickfit installs and tests software in a sandbox before it recommends it, then hands you the steps that actually worked.</p> <p align="center"> <a href="https://pickfit.app"><b>pickfit.app</b></a> · <a href="https://www.youtube.com/watch?v=Ec66x0b5u7Y"><b>Demo video (2:52)</b></a> · <a href="https://pickfit.app/agents">Connect your agent (MCP)</a> · <a href="https://ncchen99.github.io/pickfit/"><b>Full report</b></a> · <a href="docs/REPORT.md">Report (Markdown)</a> </p> <p align="center"> <img alt="Runs on Claude" src="https://img.shields.io/badge/runs%20on-Claude-1A1D17?labelColor=F5C243"> <img alt="MCP server" src="https://img.shields.io/badge/MCP-streamable%20HTTP-2E7A35"> <img alt="Google Cloud Run" src="https://img.shields.io/badge/runs%20on-Google%20Cloud%20Run-4285F4"> <img alt="License MIT" src="https://img.shields.io/badge/license-MIT-4F5149"> </p>

Why

AI agents made building software cheap, so new tools appear every day on GitHub, Show HN, Flathub and the web. Finding the right one 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.

pickfit belongs to no vendor and runs no ads or referral links. It reads each project, installs it in a fresh sandbox, checks that the features work, replays the steps offline, and only then lets it be recommended. Results are ranked by fit, not stars, and lesser-known tools that do the job well get surfaced as hidden gems.

Fitz, the mascot, is a tailorbird. Real tailorbirds pick leaves and stitch them into a nest that fits exactly. Fitz never recommends what it hasn't tried.

See it

| Ask in one sentence | Every claim comes from a run | |---|---| | <img src="brand/screenshots/home.png" alt="pickfit home page: an 'I need…' search box over a ruler, with filters for OS, CPU, free, offline and data stays local" width="420"> | <img src="brand/screenshots/tool-page.png" alt="A pickfit tool page for OpenPets: tested on macOS, a video recorded by pickfit with chapters for each checked feature" width="420"> |

As of 16 Nov 2026: 1,154 projects read, 813 verified at least through install, 715 replayed in a fresh sandbox, 98 web apps tried in a browser, 159 hidden gems. CLI tools, local AI tools, MCP servers, desktop apps (Linux and macOS) and web apps, in nine categories.

<p align="center"><a href="https://www.youtube.com/watch?v=Ec66x0b5u7Y"><img src="brand/screenshots/demo-video.jpg" alt="Watch the pickfit demo video on YouTube (2:52)" width="640"></a><br><sub>Watch the 3-minute demo: Claude Code finds and installs MangoDisk through pickfit, then the app's first use.</sub></p>

Try it

On the web: open pickfit.app and describe what you need. The site is in English and Traditional Chinese.

Made something, or found a tool that belongs here? Paste its link at pickfit.app/submit and watch it go through the same tests. Your agent can do the same with submit_software.

In your agent: pickfit is a stateless MCP server at https://pickfit.app/mcp. No account needed.

claude mcp add --transport http pickfit https://pickfit.app/mcp
{ "mcpServers": { "pickfit": { "type": "http", "url": "https://pickfit.app/mcp" } } }

Then ask in plain words: "Find me a free tool that works offline on my Mac to turn meeting recordings into subtitles, and install it."

| Tool | What it does | |---|---| | find_software | Two or three verified tools for a need and conditions (OS, CPU, free, offline), with reasons, tradeoffs and the best free option | | find_alternatives | Replacements for a tool you use now, given what you dislike and what you must keep | | get_install_guide | Only the steps that worked in pickfit's run, for your OS, plus how to check the install | | report_outcome | Did it install, did it work, are you still using it; real failures queue a fresh run | | submit_software | Send a GitHub project or a web app; it goes through the same read, install and replay |

Every answer carries how_we_know and an evidence_url, so your agent can tell you where each claim came from.

How it works

flowchart LR
  A[Discover<br>GitHub · Show HN · Flathub<br>Homebrew Cask · web · submissions] --> B[Prescreen<br>rules + Claude Haiku]
  B --> C[Read<br>Claude Opus extracts facts<br>with checked quotes]
  C --> D[Try<br>agent installs and operates it<br>Linux · macOS · Windows · browser]
  D --> E[Replay<br>fresh box, network off]
  E --> F[Score<br>quality · popularity · gem<br>security · visual check]
  F --> G[Publish<br>MCP + pickfit.app]

Since 15 Nov 2026 every product decision runs on Claude through the Anthropic API. Claude Opus 5.5 reads each README (and the project's own install guide, when the README only links to it), installs and operates the software in the sandbox, judges what passed and matches requests live; Claude Haiku 5.5 does the first cut. The sandboxes run on Modal. Each record names the model that read it, ran it and judged it.

During the hackathon (27 Oct – 4 Nov) the same jobs ran on Meta's models: Muse Spark 1.3 read READMEs, judged what passed and matched requests; Muse Spark Open (open weights, on Baseten and Modal) did the first cut and operated desktop apps from screenshots; Muse Wonder checked screenshots for visual defects. From 3 Nov those models were reached through OpenRouter, pinned to Meta as the only provider with data collection denied. Claude was then the development tool and the agent under test in the evaluation; it also wrote the tool pages' how-to text for records that existed on 2 Nov, and each such record names the model that wrote it. Records made in that period keep their model names, and nothing about them was rewritten. The full report covers that period: the hypotheses, every experiment (including the ones that failed), the iterations, and the designs that keep the catalogue honest.

Repository

| Path | What | |---|---| | pickfit/ | Python pipeline: discovery, prescreen, extraction, sandbox verification, scoring, security, publish | | worker/ | MCP server and the website. pickfit.app runs on Google Cloud Run (worker/cloudrun/); data in Cloudflare D1 and R2 | | data/software/ | The catalogue, one JSON record per tool, each with its evidence | | eval/ | Frozen evaluation questions, runner and results | | brand/ | Fitz the mascot, share cards, screenshots | | video/ | The 3-minute demo video: script, scenes, build scripts | | docs/ | The report and the working documents (index) |

Run the pipeline with uv run pickfit --help; run the Worker with cd worker && npm ci && npm test. Details in worker/README.md.

The hackathon

Built for the Meta Global AI Developer Hackathon (27 Oct – 4 Nov 2026 ET). The idea and planning notes predate the event and are not in this repository; all code, data, the website, the mascot and the video up to commit bee232a8 (5 Nov) were made during the hackathon. Rules and how we followed them: docs/hackathon-rules.md. Experiments that were tried and not merged are kept as git tags (git tag -l 'archive/*' -n3).

Work has continued since. From 15 Nov 2026 the pipeline and the live matcher use Claude, and the catalogue gained five categories (infrastructure and DevOps, security and networking, data and databases, documents and productivity, media and files) with a first batch of well-known command-line tools, MCP servers and web apps. The commit history and the hackathon's evaluation and cost records are left as they were.

License

MIT, see LICENSE. Third-party code, fonts and data are listed in NOTICE. Each tool in the catalogue keeps its own license.

Related Skills

View on GitHub
GitHub Stars3
CategoryAutomation
Updated40m ago
Forks0

Languages

Python

Trust signals

92/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

1 low