技能 / MCP：在你的代理中使用排行榜數據

呢度所有內容都係公開嘅：唔需要帳戶或 API 金鑰。將以下嘅技能檔案交畀你嘅 AI agent，或者直接呼叫 REST API。

代理技能（SKILL.md）

將此檔案儲存為 SKILL.md，並放在 skills 資料夾內，你的代理就能透過公開 API 讀取排行榜和模型能力比較。

---
name: pandanpc-leaderboard
description: Look up current AI model leaderboards and the top-10 model capability comparison from PandaNpc's public API. Use when the user asks which AI model is best, how models rank on chat, coding, image, video, speech or agent tasks, or how top models compare on benchmarks.
---

# PandaNpc leaderboard

Public, read-only JSON API. No account and no API key. Base URL: https://api.pandanpc.com/api/public/leaderboard

## Boards

GET https://api.pandanpc.com/api/public/leaderboard
GET https://api.pandanpc.com/api/public/leaderboard?category=<board key>

Response: { boards: [{ key, source, metric, meta, rows }], vendorIcons }
- With ?category=<key> the response still contains every board, but only the requested one has rows; boards with an empty rows array are shells. Always pick a board by key (for example jq '.boards[] | select(.key=="llm_chat")'), never by index.
- key: llm_chat, image_gen, code, image_edit, video, agent, vision, search, document, image_to_video, video_edit, text_to_speech, speech_to_text, speech_to_speech, or_usage, aa_intelligence
- source: arena | artificial_analysis | openrouter
- meta: { source_url, last_updated, synced_at }
- rows: [{ rank, model, vendor, score, ci, votes, ...price fields }]

Scores are comparable only inside one board. Arena boards are Elo from human blind votes; read the rank together with ci. Lower is better for speech-to-text (word error rate).

## Capability comparison

GET https://api.pandanpc.com/api/public/leaderboard/capabilities

Response: { dimensions, models, sources, fetched_at }
- models: top 10 by Epoch Capabilities Index: { rank, model, model_key, vendor, eci }
- sources: [{ source, label, license, url, updated, fetched_at, boards }] including epoch, vals, aa, google_card, browserbase, browser_use, coarena and ale
- boards: [{ key, name, capability, comparison_key, updated, fetched_at, scores }] where scores is keyed by model_key: { score, score_raw, scale, effort, harness, cost_per_task, model_version, evaluation_metadata }

Preserve every published configuration. Compare only within one source and comparison_key. Read scale: percent is 0-100, rating (Coarena) is not a percentage. Coarena is mixed-task blind human preference; show its 95% interval. ALE allows GUI/CLI tools; keep its full/max configuration, harness and run-task / split-task counts, and separate pass rate from partial score. updated is the upstream date when published; fetched_at is PandaNpc's fetch time, never a substitute source date.

## Attribution (required)

Name the source next to any number you show:
- Epoch AI, CC BY 4.0 (epoch.ai)
- Vals.ai (vals.ai)
- Artificial Analysis (artificialanalysis.ai)
- Coarena (coarena.ai) and Agents’ Last Exam, CC BY 4.0 (agents-last-exam.org), for their matching sources
- arena.ai and OpenRouter for the matching boards

Data refreshes about once a day. Cache for at least an hour.

MCP 伺服器

使用此配置將 PandaNpc MCP 伺服器新增至你的客戶端，或執行以下命令。

{
  "mcpServers": {
    "pandanpc": {
      "command": "npx",
      "args": ["-y", "@pandanpc/mcp-server"]
    }
  }
}

claude mcp add pandanpc -- npx -y @pandanpc/mcp-server

MCP 伺服器尚未提供排行榜工具。如需排行榜資料，請使用上方技能或下方 REST API。

curl 範例

# All boards (Arena, Artificial Analysis, OpenRouter) in one response
curl -s https://api.pandanpc.com/api/public/leaderboard

# One board: add ?category=<board key>, e.g. llm_chat, code, image_gen, aa_intelligence
# The response still lists every board, but only the requested one has rows (the others are empty shells)
curl -s "https://api.pandanpc.com/api/public/leaderboard?category=llm_chat"

# Top 5 models of the chat board: select the board by its key, never by position
curl -s "https://api.pandanpc.com/api/public/leaderboard?category=llm_chat" | jq '.boards[] | select(.key=="llm_chat") | .rows[:5] | map({rank, model, vendor, score})'

# Capability comparison: top 10 models by Epoch Capabilities Index x benchmarks
curl -s https://api.pandanpc.com/api/public/leaderboard/capabilities

# GPQA Diamond score of every model (Epoch AI)
curl -s https://api.pandanpc.com/api/public/leaderboard/capabilities | jq '.sources[] | select(.source=="epoch") | .boards[] | select(.key=="gpqa_diamond") | .scores'

回應欄位

領域	意思
`boards`	排行榜陣列。每個排行榜都有鍵、來源、指標、中繼資料同資料列。
`key`	排行榜鍵，例如 llm_chat、code 或 aa_intelligence。亦是能力基準測試的鍵。
`source`	資料來源：arena、artificial_analysis 或 openrouter。在能力回應中：epoch、vals 或 aa。
`metric`	這個排行榜上的分數代表甚麼，例如 Elo 或一個指數。
`meta`	source_url、last_updated（上游日期）及 synced_at（PandaNpc 上次重新整理的時間）。
`rows`	排名資料列：名次、模型、供應商、分數、置信區間、票數及可選價格欄位。
`vendorIcons`	由小寫供應商名稱對應至圖示網址的映射。
`models`	能力回應：按 Epoch 能力指數排名前 10 的模型，每項包含排名、模型、model_key、供應商及 eci。
`sources`	能力回應：每個發佈者各設一項，包括 Coarena 及 ALE，並各自擁有其排行榜與擷取時間。來源及比較群組絕不混合。
`scores`	每項基準測試及模型：score、score_raw、scale、effort、harness、model_version 及 evaluation_metadata。百分比分數採用 0–100；Coarena 偏好評分則不採用。保留置信區間及任務數量。
`updated`	上游基準日期（如有公佈）；null 表示未公佈。fetched_at 則另行記錄 PandaNpc 擷取來源的時間。

歸因

資料來自 Epoch AI（CC BY 4.0）、Vals.ai、Artificial Analysis（artificialanalysis.ai）、arena.ai 及 OpenRouter。顯示任何數字時，請將來源名稱保留在數字旁邊。