Agent-Leaderboard

Agent Leaderboard

Discover GitHub repositories across Agent Skills, MCP Servers, Prompt Libraries, AI Frameworks, and Auto Research. Browse by total stars, find projects gaining momentum, or search the same catalog from Chrome and an MCP client.

中文文档 · Browse the boards · Chrome extension

Five boards

Board	What it covers
Agent Skills	Skills and tools for Claude Code, Codex, Cursor, Copilot, Gemini, DeepSeek, OpenClaw, Hermes Agent and more
MCP Servers	Model Context Protocol servers for databases, files, web tools, APIs and development workflows
Prompt Libraries	Prompt collections, system prompts and prompt engineering resources
AI Frameworks	Agent orchestration, multi-agent systems, memory and model routing
Auto Research	Deep research, literature discovery, data research and knowledge tools

The website shows each board’s repository count, data timestamp and ranking. Use the board links for current results; this README does not duplicate a manually maintained Top 10 snapshot. Boards can overlap, so their counts are not counts of unique repositories across the site.

Trending

Click the highlighted Trending button in the top navigation to open the ranking on the current page. On mobile, it is inside the navigation menu. The entry briefly wiggles on its first appearance, unless reduced motion is enabled.

• Daily + growth rate by default. Switch to Weekly or Monthly, or sort by Stars gained.
• Top 20 per board. All five boards use the same controls; no coding-agent subcategory filter is applied inside Trending.
• Compact rows. Project and category icons, 🥇🥈🥉 medals for the top three, total stars, gains and percentages stay easy to compare. Large star counts use k; hover for exact values.
• Minimize and resume. Minimize to the bottom-right dock and restore the selected board, period, sort and scroll position. Close with ×, Escape or a click outside the panel. Homepage filters and scroll position stay intact.

Growth is calculated from the site’s recorded daily snapshots: stars gained = current − baseline, and growth rate = stars gained ÷ baseline × 100%. Daily, weekly and monthly compare UTC dates 1, 7 and 30 days apart, relative to each dataset’s update date. Only positive gains rank; missing history is excluded, and a zero baseline is excluded from growth-rate ranking. The summary’s net gain includes declines.

These are observed changes in the catalog, not real-time GitHub star events: incremental collection can retain older counts. See trend data and calculation details.

Website features

Feature	Behavior
Search and filters	Repository name/description, category, language, minimum stars and last-update time range
Use cases	Multiple selected use-case tags must all match
Grid and list	Switch views; display 24, 48 or 96 repositories per page
Favorites	Local browser storage; optional GitHub sign-in and account favorites through a Cloudflare Worker and D1
Language and theme	Chinese/English; follows system light/dark theme until manually changed
Brand identification	Reuses brand and category icons across the homepage and trend panel

GitHub sign-in and cloud favorites

The website uses a small Cloudflare Worker on a separate workers.dev address and a D1 database for GitHub sign-in and account favorites. Existing local favorites are merged on first sign-in; signed-in additions and removals save through the online API. Failed requests restore the previous state and show a retry message. Signing out or clearing browser data keeps successfully saved cloud favorites. GitHub Stars are unchanged. The v1.7.5 Chrome extension automatically syncs the same account’s favorites from an already signed-in website page, without navigation or confirmation. Website and extension sessions stay separate.

Local favorites work without the API. Real sign-in requires a Cloudflare account, a GitHub App or OAuth App, D1 migration and Worker deployment. DNS can remain with Tencent Cloud and the static site with GitHub Pages. GitHub credentials stay in Worker secrets; the browser persists its own site credential in localStorage for a fixed 30-day login, surviving tab closure and browser restart. OAuth verification stays in same-tab sessionStorage. No GitHub token or site token appears in URLs. Explicit sign-out or clearing site data removes the local login. See deployment, privacy and verification steps. Code and mocked tests alone do not activate the production service.

Chrome search extension

Download the v1.7.5 ZIP with trends and cloud favorites and load it unpacked, or visit the bilingual showcase and usage guide. The Chrome Web Store currently provides v1.6; the v1.7.5 store update requires review. Requires Chrome 116 or later.

Search all five boards in a side panel with categories, use cases, platform, language, stars and time filters. Save favorites, open READMEs, and copy an AI installation prompt to paste into your coding assistant. The extension generates the prompt; it does not install repositories automatically.

Choose Trending for the same daily, weekly and monthly top 20 as the website, ranked by growth rate or stars gained across all five boards. The extension includes an offline trend snapshot, shows its observation dates, and updates it alongside the catalog. These are recorded changes, not live GitHub events; missing history is shown explicitly.

It also supports selected-text context-menu search, the agent address-bar keyword, Ctrl+Shift+Y / macOS Command+Shift+Y, and a bundled offline catalog. Anonymous favorites stay local to this Chrome profile. In v1.7.5, opening the extension or Favorites syncs directly from an already signed-in website page, without moving or navigating it. If not signed in, “Sync website favorites” opens the main website’s Favorites dialog. Completing GitHub sign-in syncs automatically. Refresh older website tabs once after upgrading. Existing local favorites merge into the same account shared with the website and other devices. Signed-in saves require the network; failed changes roll back. The extension persists its own site credential for up to 30 days in trusted-context-only chrome.storage.local, surviving browser restart and extension updates. Account caches and PKCE proofs stay in chrome.storage.session. Explicit sign-out disables automatic login and preserves cloud favorites. Existing eight-hour sessions keep their original deadline; signing in again after expiry uses the new lifetime.

For development, load chrome/ through chrome://extensions → Developer mode → Load unpacked. Refresh the bundled catalog and build dist/Chrome.zip with:

python3 scripts/chrome/build.py

See the extension documentation for permissions, updates, shortcuts and testing.

MCP skill discovery

Use the catalog directly from any AI chat client that supports MCP. This Bash command installs the server, creates an isolated venv under ~/.local/share/skills-discovery-mcp, and configures detected clients:

curl -fsSL https://raw.githubusercontent.com/jaychempan/Agent-Leaderboard/main/scripts/install.sh \
  | SKILLS_DISCOVERY_CONFIGURE_CLIENTS=auto bash

Use macOS, Linux, WSL, or Git Bash for the installer. Native PowerShell does not support this Bash syntax.

For Codex, verify the server is registered, then start a new Codex session:

codex mcp list

Quick examples are in docs/mcp-quick-use.zh.md.

To skip client config or target specific clients:

SKILLS_DISCOVERY_CONFIGURE_CLIENTS=none bash scripts/install.sh
SKILLS_DISCOVERY_CONFIGURE_CLIENTS=codex,claude,cursor bash scripts/install.sh

The installer backs up config files before editing them. If your client is not configured automatically, add this fallback manually:

{
  "mcpServers": {
    "skills-discovery": {
      "command": "skills-discovery-mcp"
    }
  }
}

Re-run scripts/install.sh to update an existing install. To uninstall:

curl -fsSL https://raw.githubusercontent.com/jaychempan/Agent-Leaderboard/main/scripts/uninstall.sh | bash

You can also run the server directly from a cloned checkout:

python3 -m mcp.skills_discovery.server

By default the server reads the daily remote catalog at https://agentskills.media/data/discovery_index.json, generated from the leaderboard data. To test a local or custom index, point it at an absolute file URL:

SKILLS_DISCOVERY_INDEX_URL=file:///absolute/path/to/data/discovery_index.json \
python3 -m mcp.skills_discovery.server

Example chat queries:

• Find Codex testing skills
• Recommend Claude UI/UX skills
• Top MCP browser automation servers

The server returns guidance, matching repositories, and links only. It does not execute install commands or modify your environment.

Run locally

The checkout includes data, so you can preview the website without fetching GitHub or installing frontend dependencies. Python 3.12 is recommended and is used by the daily workflow.

git clone https://github.com/jaychempan/Agent-Leaderboard.git
cd Agent-Leaderboard
python3 scripts/serve.py --port 8080

Open http://127.0.0.1:8080. The website is plain HTML, CSS and JavaScript with no frontend build step.

Visit statistics use the public soxft/busuanzi API at https://busuanzi.9420.ltd/api, compatible with static GitHub Pages. The homepage displays daily page views (PV) and total visitors (UV). Both https://agentskills.media and https://jaychempan.github.io/Agent-Leaderboard/ use the canonical agentskills.media namespace. Each document sends one POST; board, filter and language changes do not recount. A virtual /__visits/YYYY-MM-DD path groups loads by Asia/Shanghai date, derived from the browser clock. Its page_pv is the daily count; site_uv is the service’s cumulative deduplicated visitor count, not the sum of daily UV. Reload to update the snapshot. A new date selects a new daily bucket without subtracting cumulative PV. Correct device time is required.

UV estimates visitors using IP, User-Agent and a signed anonymous identity stored as st_visit_identity_v1 in localStorage; it does not identify real people. Clearing storage or changing devices can affect deduplication. Requests send only the canonical domain, statistics date and anonymous identity, without searches, query strings, referring-page URLs or third-party cookies. The public service has no SLA or data-integrity guarantee; for controlled retention, self-host the same open-source service and change VISIT_API. The former busuanzi.cc endpoint timed out (524) during verification and has been removed. Its history is not migrated: counts start with the new integration, without estimated offsets. Failures show an unavailable state; missing individual metrics show —, never fake zero. Ambiguous POST failures are not automatically retried.

Run python3 scripts/serve.py for isolated local daily PV and anonymous browser UV through /api/visits, persisted in .local/visits.sqlite3. Existing local PV history is retained; local UV starts with this feature. Plain python3 -m http.server does not provide the API. Check with node --test tests/visit-counter.test.cjs and python3 -m unittest discover -s tests -p 'test_preview_server.py'. Browser integration checks are in tests/visit-counter-browser.cjs (mocked by default; --live uses the real service). The deploy workflow checks for conflicting or duplicate homepage scripts before publishing.

Refresh data

Run from the repository root. Fetchers use the Python standard library. A GitHub token is optional locally; the scripts accept GITHUB_TOKEN / GH_TOKEN from the environment or a --token argument.

python3 scripts/fetch_data.py
python3 scripts/fetch_mcp.py
python3 scripts/fetch_prompts.py
python3 scripts/fetch_frameworks.py
python3 scripts/fetch_auto_research.py
python3 scripts/build_discovery_index.py
python3 scripts/build_trends.py

The discovery builder produces the catalog consumed by MCP and the extension. The trend builder records daily snapshots, retains up to 40 days and writes compact browser data. For a one-time recovery from available Git history, use python3 scripts/build_trends.py --from-git; this needs the relevant historical commits. Missing baseline dates remain unavailable rather than being replaced by shorter periods.

Validate changes

python3 -m unittest discover -s tests -p 'test_*.py'
node --test tests/chrome/*.test.mjs

For trend interaction and responsive layout checks, install playwright-core, keep the local server running, and supply your Chrome executable:

SITE_BASE_URL=http://127.0.0.1:8080/ \
CHROME_PATH=/absolute/path/to/chrome \
node tests/trends-browser.cjs

If needed, set PLAYWRIGHT_MODULE to an absolute path to the playwright-core module. Browser screenshots are written to the ignored artifacts/trends/ directory.

Automated updates and deployment

Workflow	Trigger	What it does
Daily Data Update	Daily at 02:00 UTC / 10:00 China time; manual dispatch	Refreshes all five boards, rebuilds discovery and trend data, and commits changed outputs
Deploy to GitHub Pages	Push to main, completion of the daily workflow, or manual dispatch	Builds the extension download and deploys the static site

The daily workflow passes the built-in github.token as GH_TOKEN and requests contents: write; a separately configured GH_TOKEN secret is not required by the current workflow. For a fork, enable Actions and select GitHub Actions as the Pages source. Use Actions → Daily Data Update → Run workflow for a manual data refresh.

Data coverage

All five collectors currently use a 50-star threshold. Repository names, descriptions and topics drive discovery and classification; relevance rules and exclusions vary by board. Categories and platform labels are discovery aids, not guarantees of compatibility or quality. The data comes from GitHub search and incremental merges, so it is not an exhaustive directory or a guarantee that every repository is refreshed on each run.

Repository structure

Agent-Leaderboard/
├── index.html                   # Five boards and the trending overlay
├── src/
│   ├── shared.js / shared.css   # Routing, filters, themes and shared UI
│   └── trends.js / trends.css   # Trending rankings and panel interaction
├── data/
│   ├── *.json / *.js            # Board datasets and discovery index
│   ├── star_history.json       # Daily snapshots; not loaded by the browser
│   └── trends.json / trends.js  # Compact 1 / 7 / 30-day star changes
├── scripts/
│   ├── fetch_*.py               # GitHub collection and shared helpers
│   ├── build_discovery_index.py # Catalog for MCP and the extension
│   ├── build_trends.py          # Daily history and trend generation
│   ├── install.sh / uninstall.sh
│   └── chrome/                 # Extension packaging and assets
├── mcp/skills_discovery/        # Local MCP discovery server
├── chrome/                     # Loadable Chrome extension
├── extension/                  # Extension showcase and guides
├── trending/index.html         # Legacy URL redirect to the overlay
├── docs/                       # Usage and data documentation
├── tests/                      # Python, extension and browser checks
└── .github/workflows/          # Daily data refresh and Pages deployment

Contributing

Improvements to discovery queries, classification, documentation and the UI are welcome. Keep the English and Chinese READMEs aligned, and run the checks relevant to your change.

Licensed under MIT.