Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Repo Finder

    2

    Finds the best open-source GitHub projects for your need and checks their health before you depend on one.

    Free

    Security scanned
    Repo Finder

    Repo Finder

    Example session with this skill installed

    /find-repo self-hosted bookmark manager that archives pages

    • Read your context and instructions
    • Compiled the repo finder

    I found about 1,250 candidates and picked these. The first six are the closest fit: bookmark managers that save a copy of each page.

    Bookmark managers that archive pages

    1. linkwarden/linkwarden has 20.3k stars and was last pushed 2026-10.
      It saves each link as a screenshot, a PDF and a readable copy, and supports collections and sharing. It's the most archive-focused of the group. It's AGPL-3.0 and heavier to run (Postgres plus a browser for capture).
    2. karakeep-app/karakeep has 30.2k stars and was last pushed 2026-11.
      It's the project formerly called Hoarder. It bookmarks links, notes and images, keeps a full-page archive and offers AI tagging and search.
      It's AGPL-3.0, and the AI features are optional.
    3. sissbruecker/linkding has 11.4k stars and was last pushed 2026-10.
      It's minimal and fast, and it stores local HTML snapshots through a SingleFile option or the Internet Archive.
      It's MIT, and archiving is a setting you turn on rather than the main feature.
    4. go-shiori/shiori has 11.7k stars and was last pushed 2026-10.
      It's a single Go binary that stores an offline readable copy of each page.
      It's MIT, simple, and has few features beyond that.
    5. asciimoo/omnom has 666 stars and was last pushed 2026-10.
      It preserves web content through a browser extension that captures the page as you see it.
      It's AGPL-3.0, and smaller and more niche than the others.
    6. Kovah/LinkAce has 3.4k stars and was last pushed 2026-10.
      It's a PHP link archive with tags and lists. Its archiving sends pages to the Internet Archive, so it keeps no local copy. It's GPL-3.0.

    Archivers and read-it-later apps

    1. ArchiveBox/ArchiveBox has 28.9k stars and was last pushed 2026-11.
      It's a self-hosted web archiver that saves each page in many formats (HTML, PDF, WARC, screenshot).
      It's MIT. It's an archiver first and a bookmark manager second, so pick it if preservation matters most.
    2. wallabag/wallabag has 13.0k stars and was last pushed 2026-10.
      It saves articles as clean readable text.
      It's MIT, and it suits articles better than general bookmarks.
    3. kanishka-linux/reminiscence has 1.9k stars and was last pushed 2024-04.
      It's described as a self-hosted bookmark and archive manager, which matches your wording closely. I'd skip it because it has had no updates for over two years.

    I left out Shaarli, Servas, Grimoire and the start-page dashboards. They manage links but don't archive pages. Omnivore is also left out because I believe it has shut down; I didn't check.

    Which to use

    If you want to run something rather than build it, start with Linkwarden or Karakeep. Both archive pages by default and are actively developed. Pick Linkding if you want something light on resources, or ArchiveBox if preserving pages matters more than organizing bookmarks. I'd advise against building your own, because the hard part (reliable page capture) is already solved in these projects.

    Want me to run the check-repo skill on Linkwarden, Karakeep and Linkding? It would look at maintainers, release cadence, issue backlog and licence.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Your agent is about to build something that already exists. Or it recommends a project from memory: an old name, an abandoned repo, the wrong platform.

    GitHub's own search matches words, not needs. Ask for "an app that converts PDF to Word" and you get Word-to-PDF tools, image converters and 1-star experiments, while the best projects describe themselves in other words ("A black hole for Internet advertisements" is Pi-hole). Asking the model instead gets you what it remembers: in our tests Sonnet's memory alone gave stale repository names and wrong-platform picks.

    Repo Finder gives your agent a real search: about 300 candidates gathered from about 20 GitHub searches and ranked without any model, in about 4k tokens. Your agent then reads all 300 and gives you only the best 5-10, each with why it fits.

    How the search works

    Describe the need in plain words, for example "self-hosted bookmark manager". The script then:

    1. Searches GitHub about a dozen ways at once. It tries phrasings of your need ("bookmark manager" as a phrase, "bookmark" by stars, "bookmark self-hosted"), plus the field's own words and GitHub topics your agent adds ("archiver", "wayback", read-it-later), because projects rarely describe themselves in your everyday words. Each search returns up to 100 repositories, about 1,400 different ones in all.
    2. Follows the topics the best matches share. Projects that describe themselves in other words turn up here: Scrapy for "scrape", Logseq for "an alternative to Notion".
    3. Drops the noise. Curated lists, tutorials, courses, homework and dotfiles go, unless you asked for them.
    4. Ranks without AI models. A repository ranks higher when it uses your words, uses the words its best-matching neighbours use, comes up in several searches, and is popular.
    5. Looks once more for niche projects. It finds topics that are common among the 50 best repositories but rare in the rest, and fetches the most-starred projects under them. That is how bandwhich turns up for "see which process uses my network bandwidth", which no word search returned.
    6. Prints a compact list. The top 60 get a full line (stars, language, licence, last push, description, topics). The next 240 get only their name and stars: famous projects with vague descriptions are recognised by name.

    Your agent then reads all 300 lines, drops look-alikes, adds well-known projects GitHub missed, and answers with the best 5-10: each with stars, last push, why it fits and any caveat (licence, inactive, platform).

    Two skills, one package

    find-repo runs the search above, then tells the agent how to judge it:

    • keep famous projects wherever they are in the list
    • drop look-alikes, such as converters that go the wrong direction
    • for "an alternative to X", leave X out
    • add well-known projects GitHub missed, and confirm them with check-repo
    • if you were about to build it, say whether to depend on the best one, fork it, borrow its approach or build anyway

    check-repo checks up to five repositories in one GitHub call and compares them:

    • Activity: commits in the last 90 days, last push, latest release
    • Community: people with push access, open and closed issues, open and merged pull requests
    • Profile: stars, forks, language, licence, creation date, website, topics and the start of the README
    • Warning flags: archived, a fork, no licence or a copyleft one, no push for a year or no commits in 90 days, one or two people with push access, a backlog (over 50 open issues and more open than closed), no releases

    Flags

    Your agent adds these when your request calls for them; you can also ask for them by name.

    find-repo

    FlagWhat it does
    --also archiver waybackUp to 6 of the field's own words, one per search, that your agent adds (it always should); they also count at half weight in the ranking
    --topics read-it-laterUp to 6 GitHub topics your agent adds
    --language GoOnly repositories in that language; javascript also matches TypeScript. A language named in the need ("a Rust library for...") is picked up without it
    --license MITOnly that licence (SPDX id); repeat for several: --license MIT --license Apache-2.0
    --min-stars 50At least that many stars
    --pushed-within 365Pushed in the last that many days, to skip abandoned projects
    --smallNo popularity bonus, for "little-known" or "small" projects
    --no-refineSkips step 5: about 6 seconds faster, fewer niche finds
    --limit 60How many candidates get full lines (about 35 tokens each)
    --names 240How many more follow as names and stars only (about 7 tokens each)

    check-repo

    ArgumentWhat it does
    owner/name ...Up to five repositories, as owner/name or GitHub URLs, in one GitHub call
    --readme 1500Characters of each README to show; --readme 0 leaves it out (about 400 tokens less per repository)

    Who it is for

    • Developers who want to know what already exists before writing a library, CLI or app
    • Anyone asking an agent for "an open-source alternative to X" or "a self-hosted tool for Y"
    • Teams choosing between dependencies, who need maintenance and licence facts rather than impressions

    How it is built

    Two Python scripts, standard library only, about 32 KB together, typed and documented. Each skill is one SKILL.md that loads only its own script. The ranking uses classic search techniques instead of AI models: BM25 word matching, pseudo-relevance feedback (learning words from the best matches) and reciprocal rank fusion (combining the order of many searches).

    Works with zero installs beyond Python 3.8+ and the GitHub CLI, logged in (gh auth login). No models, no downloads, no server. The scripts call only api.github.com through gh, read no environment variables or files and write nothing.

    Results

    Measured on 50 everyday needs and 10 more niche ones. Below is a comparison of the first version and the current one.

    • Right repos in the list, 50 tests: 72% -> 84% (90% of the must-haves)
    • Right repos in the list, 10 niche tests: <60% -> 88% (90% of the must-haves)
    • Right repos in the top 60 lines alone: 68% (50 tests) and 82% (10 niche tests)
    • List size: ~12k tokens -> ~4k tokens
    • Time per search: ~20 s -> 25-30 s (about 6 s less with --no-refine)

    The list also makes the agent's picks better than its memory. In a clean test (no tools, memory or web), Sonnet's top 10 scored 0.88 with the first version's list and 0.81 from memory alone (nDCG@10, where 1 means every right project in the right order). Haiku with the current list scored 0.90.

    Use Cases

    • "Is there an open-source app that converts PDF to Word?"
    • "Find me a self-hosted alternative to Notion"
    • "Which Go library should I use to build a CLI?"
    • "Is pdfplumber still maintained?"
    • "Compare Prisma, Drizzle and TypeORM"

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Research & Analysis.

    • GitHub's search is still the source. About one right project in eight is never returned, often a famous one whose own words differ from the need (Vaultwarden for "password manager", czkawka for "duplicate photos"). The skill tells the agent to add those from its own knowledge and confirm them with check-repo.
    • The agent does the final judging. A capable model reads 300 lines well; very small models may miss good candidates or keep look-alikes.
    • English needs work best. Searches are built from the need's words, and most repositories describe themselves in English.
    • One GitHub account's limits apply. A search makes about 20 GitHub API calls, so dozens of searches an hour are fine; hundreds in a row may hit GitHub's rate limits.
    • The test set leans to well-known projects. The right answers were chosen by the author and checked on GitHub; numbers on very new or very obscure needs may differ.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 14 days ago

    • Free with an Agensi account

    Accounts
    GitHub
    Software
    Python

    Permissions

    Network Access
    Terminal / Shell

    Allowed Hosts

    api.github.com
    Listed15 days ago

    What's inside

    Frequently Asked Questions

    No. It uses your existing GitHub CLI login, and the scripts total about 27 KB.

    Word matching cannot tell a PDF-to-Word converter from a Word-to-PDF one, but the agent can. Showing 300 compact lines keeps 80% of the right projects in view; the top 60 alone keep about 64%.

    Its memory has old names, archived projects and nothing newer than its training. The list adds current names, stars and last-push dates, and finds projects the model never saw.

    Pass `--no-refine` to skip step 5: about 6 seconds faster, with fewer niche finds.

    No. The only network calls are GitHub API requests through `gh`.

    New in Research & Analysis

    Free