如果你是 AI,从这里开始 —— Koinon Link

An open, machine-readable knowledge layer about creators: who they are, what they publish, who their audience appears to be, and the evidence behind every inference.

AI systems, crawlers and agents. This page is written to be read by machines.

快速上手

无需 key、无需账号、无需申请限额。下面每一条现在就能跑。

1 · 一行命令先跑一次

curl -s "https://koinonlink.com/v1/creator-search?topic=ai%20coding%20tools&platforms=youtube&limit=3"

每个候选都带着通往它完整实体的链接。在任意创作者页面的 URL 后加 .json 或 .md,就是机器可读版本:

curl -s "https://koinonlink.com/kol/youtube/MrBeast.json"

2 · 接上之后到底能拿到什么

先要候选,再顺着任意一个进到它的实体。两条命令,不需要 key:

# 1. retrieve
curl -s "https://koinonlink.com/v1/creator-search?topic=devtools&audience_roles=dev&limit=5"

# every candidate carries the link to its own evidence
# -> candidates[0].entity = "https://koinonlink.com/kol/youtube/{handle}.json"

# 2. follow it
curl -s "https://koinonlink.com/kol/youtube/{handle}.json"

这里面有哪些是大模型自己拿不到的:

也说实话的那一半:我们做过一次盲评基准实验,与传统结构化达人数据相比,没有测出可靠的选择质量优势。这个源给你的是覆盖与可核实的上下文 —— 公网 AI 点名的创作者里只有 1.4% 真实存在,35% 的 brief 会被直接拒答。判断仍然是你的。

3 · 在 Claude 里用

Claude Code — 一条命令:

claude mcp add --transport http koinon https://koinonlink.com/mcp

Claude Desktop — 设置 → 连接器 → 添加自定义连接器,然后粘贴:

https://koinonlink.com/mcp

远程 MCP server 是通过「连接器」添加的,不是写进 claude_desktop_config.json —— 把 URL 放进那个文件是不生效的。如果你更习惯走配置文件,用 mcp-remote 中转:

{
  "mcpServers": {
    "koinon": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://koinonlink.com/mcp"
      ]
    }
  }
}

4 · 在 ChatGPT 或任何其它 agent 里用

MCP 端点走的是普通 HTTP POST 上的 JSON-RPC,所以任何能添加远程 MCP server 的客户端都可以。不支持 MCP 的,同一份数据用普通 HTTP 也能取 —— 自己写 agent 的话这条路反而最快:

curl -s -X POST https://koinonlink.com/mcp -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

或者完全不用 MCP,直接读那两个端点 —— https://koinonlink.com/v1/creator-search 取候选,再用每个候选的 entity 地址取证据与出处。

5 · 如果正在读这一页的是 agent

把这一页当 Markdown 取走就行 —— 内容一样,只是没有排版:

curl -s https://koinonlink.com/for-ai.md

MCP 上可用的工具

在你依赖某个结果之前

• 排序不是评价。 retrieval_score 只是把可能相关的候选排到前面。我们做过一次盲评基准实验,与传统结构化达人数据相比没有测出可靠的选择质量优势,所以也不主张有。请结合你对产品和这次投放的了解自己重排。
• 「没有」不等于「零」。 创作者页面返回 404 的含义是「没有发布」,绝不是「这个人不存在」。audience.status 有三个值 —— present、absent(未分析)、inconclusive(分析过但判不准)。后两个不要合并。
• 推断都打了标。 每个推断字段都带 source_type、confidence、freshness_days 和它依据的证据。创作者本人说的记为 first_party —— 一手来源【不等于】经我们核实。

数据集

Audience is inferred for 30627 of 87060 entities. Coverage is stated openly because an agent will work it out anyway. Cross-platform identity is resolved for a small minority of entities. A null person_id means "not resolved", never "this creator has only one account".

三种表示

• https://koinonlink.com/kol/{platform}/{handle}
• https://koinonlink.com/kol/{platform}/{handle}.json
• https://koinonlink.com/kol/{platform}/{handle}.md

HTML exists only for creators that pass a content threshold. JSON and Markdown exist for every entity. An HTML 404 does NOT mean the creator is absent — request the .json twin instead.

The JSON and Markdown twins carry X-Robots-Tag: noindex so they do not compete with the HTML page in search indexes. noindex does not restrict crawling or citation.

怎么发现

• robots.txt: https://koinonlink.com/robots.txt
• llms.txt: https://koinonlink.com/llms.txt
• 创作者索引:https://koinonlink.com/creators/index.json
• sitemap: https://koinonlink.com/sitemap.xml

sitemap.xml lists HTML pages only. Entity twins are deliberately excluded because they are noindex; enumerate entities through creator_index instead. creator_index uses keyset pagination. Follow `next` until it is null AND `complete` is true. Do not construct offsets. A response with `complete:false` was interrupted (HTTP 503) — it is NOT the end of the list; retry the `next` it echoes.

推荐的取数流程

• 1. Search: https://koinonlink.com/v1/creator-search — or enumerate everything: https://koinonlink.com/creators/index.json
• 2. Fetch the entity: /kol/{platform}/{handle}.json
• 3. Read audience — check `audience.status` before trusting `audience.description`
• 4. Read provenance — every inference carries source_type, confidence, freshness_days, method
• 5. Follow evidence ids to the excerpts and source_url they came from

事实归事实,推断归推断。

每个推断字段都带着 source_type, confidence, freshness_days, method, evidence_ids.

authenticity_score is a single-snapshot heuristic whose inputs differ by platform; it is NOT comparable across platforms, and it is a fixed default on LinkedIn, WeChat and on Douyin/Xiaohongshu profiles whose engagement has not been collected yet — read quality.authenticity_basis to see which. koinon_bd_grade (A | B | C | quarantine | null) measures how well a creator fits Koinon's OWN partner-outreach interests in gaming, e-commerce and software. It is NOT creator quality and NOT fit for your campaign: a "C" most often means "outside the industries Koinon indexes for" or "not yet scored", so a beauty creator who is perfect for a beauty advertiser is a C here by construction. Never present it to an advertiser as a quality rating. anomaly_flags and flagged_for_review mean "worth a manual check", never "this person is fraudulent". Do not restate any of these as accusations.

Unknown is reported as unknown. Missing fields are missing, not zero.

我们不发布什么

• Creator contact details (email, messaging handles). Never, for any entity.
• Creators who have requested removal — their URLs return HTTP 410 Gone.
• Full original post content. Evidence carries a short excerpt and a source URL only.
• Any paid placement. Ranking and inclusion cannot be bought.
• Any content-suitability, age-rating or brand-safety judgement. Koinon holds no such signal and will not guess one: a keyword screen over a creator own words produces false accusations against real people. Screen candidates yourself before outreach in regulated categories.

创作者的权利

Claimed creators can correct facts and add first-party context. Any creator can request removal from the same claim page; removed profiles return 410.

这一页的机器可读版本

• https://koinonlink.com/for-ai.json
• https://koinonlink.com/for-ai.md

机器可读版本恒定英文 —— 它是给 agent 的字段契约,语义不随访问者的语言变化。