BusyBee MCP

Public social media data for AI agents. Ten business-level tools, one schema — TikTok, Instagram, YouTube, X (Twitter), Douyin and Xiaohongshu.

Ask your agent "who are the fastest-growing coffee creators on TikTok right now, and what do their comment sections complain about?" — it answers with real data, in one conversation.

Endpoint    https://lucydomi.com/mcp
Protocols   MCP Streamable HTTP · MCP SSE (legacy) · REST/JSON · stdio
Tools       10
Billing     pay per call; every response returns cost_usd + balance_usd

Try it in 30 seconds — no signup: open https://lucydomi.com/demo, type a keyword, see the real response shape (6 free calls per device per day, 3 of them on Douyin/Xiaohongshu). When you want unlimited calls, grab a key at https://lucydomi.com/sign-up.

Connect your agent

Step-by-step guides, each with a config file you can copy verbatim:

• Claude Desktop / Claude Code — local stdio via uvx
• Cursor / Cline / Windsurf / VS Code — remote HTTP with a bearer header
• Dify / n8n / Coze / FastGPT — import the OpenAPI spec, no code
• Plain HTTP / curl / any language — REST endpoints and the tools JSON

Works with every kind of agent

Most "social data for AI" products ship one transport and hope your client speaks it. This server speaks four, on the same host, with the same 10 tools behind all of them.

Discovery document (machine-readable, tells an agent what's available): https://lucydomi.com/mcp/info

The 10 tools

get_comment_insights is the one people keep open: it samples comments, then returns sentiment split, question ratio, high-frequency phrases and top-liked comments — so the agent answers "what do users actually want" instead of dumping fifty raw rows.

Docs by platform, by tool, by client

Each page carries a real captured response and copy-paste code, so you can skip the README once you know what you need:

Tutorials — end-to-end workflows with real numbers: mine demand from comments · build a creator database · cross-platform topic check · competitor monitoring

All 21 in one index: https://lucydomi.com/resources · machine index: https://lucydomi.com/llms.txt · everything in one file for LLMs: https://lucydomi.com/llms-full.txt

Cookbook — three scripts you can run right now

Stdlib only, no dependencies. Each was run against the live API before being committed; the output below is real.

export BUSYBEE_API_KEY=sk-...
python examples/cookbook_cross_platform.py coffee

平台          条数    最高互动      文案
tiktok      10    9325976   ASMR SATISFACTION 🤎✨ Nothing beats puck …
youtube     10    235306    The Espresso Revolution is Already Happe…
twitter     10    20988483  NEW: Bikini barista takes out a hammer o…
instagram   -     -         (该平台没有关键词搜内容)
douyin      10    75063     coffee time.#海氏C9Pro咖啡机…
xiaohongshu 10    14623     谁发明的这样喝咖啡？…

Note the Instagram row: it reports the gap instead of silently returning nothing — list_platforms tells your agent which tools each platform actually supports, so cross-platform jobs degrade instead of breaking.

One schema, every platform

Same field names whether the item came from TikTok, Instagram, YouTube, X, Douyin or Xiaohongshu:

{
  "ok": true, "platform": "tiktok", "tool": "search_posts", "count": 6,
  "items": [
    {
      "platform": "tiktok",
      "id": "7690199760430566669",
      "url": "https://www.tiktok.com/@laceyglasley/video/7690199760430566669",
      "author": { "id": "7066058701889012782", "handle": "laceyglasley", "name": "Lacey" },
      "text": "Salted Honey Spanish Latte … #coffeetok #spanishlatte",
      "created_at": "2026-09-27T13:03:24+00:00",
      "stats": { "likes": 68357, "comments": 165, "shares": 7691, "views": 1009685, "collects": 17806 },
      "cover": "https://…"
    }
  ],
  "cost_usd": 0.003, "balance_usd": 3.93
}

This is a real response, not a mock-up — captured from a live call, only the share URL query string and the caption were trimmed. Two full raw responses are checked in: examples/1_tiktok_coffee.json (search_posts) and examples/2_youtube_insights.json (comment insights on 20 comments).

Every response carries cost_usd and balance_usd, so the agent can budget itself and you can see what a task actually cost.

Setup per client

Need a key first: register at lucydomi.com → Console → API Keys → New key (sk-…). Overseas and China platforms use different key groups.

Claude Code

claude mcp add busybee --transport http https://lucydomi.com/mcp \
  --header "Authorization: Bearer sk-YOUR_KEY"

Cursor / Cline / Windsurf / VS Code (~/.cursor/mcp.json)

{
  "mcpServers": {
    "busybee": {
      "url": "https://lucydomi.com/mcp",
      "headers": { "Authorization": "Bearer sk-YOUR_KEY" }
    }
  }
}

Claude Desktop (local stdio)

{
  "mcpServers": {
    "busybee": {
      "command": "uvx",
      "args": ["busybee-mcp"],
      "args": ["busybee-mcp"],
      "env": { "BUSYBEE_API_KEY": "sk-YOUR_KEY" }
    }
  }
}

Legacy SSE clients

SSE URL      https://lucydomi.com/sse
Messages URL https://lucydomi.com/messages/?session_id=<from the SSE event>
Header       Authorization: Bearer sk-YOUR_KEY

Dify / n8n / Coze / GPT Actions

Import the OpenAPI schema, set auth to Bearer sk-YOUR_KEY. If the importer rejects 3.1, use the 3.0 view.

Server-side endpoints derive automatically from the schema; each operation maps 1:1 to a tool name, so search_posts in the schema is POST https://lucydomi.com/mcp/v1/search_posts.

LangChain / custom Python (no MCP needed)

import json, urllib.request
req = urllib.request.Request(
    "https://lucydomi.com/mcp/v1/search_creators",
    data=json.dumps({"platform": "tiktok", "keyword": "coffee", "limit": 3}).encode(),
    headers={"Content-Type": "application/json", "Authorization": "Bearer sk-YOUR_KEY"},
)
print(json.load(urllib.request.urlopen(req)))

Tool schemas for function calling: GET https://lucydomi.com/mcp/tools?format=openai

Local package

pip install busybee-mcp
BUSYBEE_API_KEY=sk-YOUR_KEY busybee-mcp          # stdio
BUSYBEE_API_KEY=sk-YOUR_KEY busybee-mcp http     # serve all transports on :3013

Design decisions worth knowing

• Ten tools, not a thousand. Call a business intent, not an endpoint. The more tools a model sees, the more often it picks wrong and burns money.
• Normalized fields. author / text / created_at / stats mean the same thing on every platform.
• One implementation behind every protocol. MCP, SSE and REST share the same dispatch layer, so behavior can't drift between clients.
• Pay per call, priced per endpoint. Cheap calls cost fractions of a cent. Price is visible in your console.
• Cost transparency inside the result. cost_usd + balance_usd on every response.
• Failures are not billed.
• Clean errors. No stack traces, no third-party vendor names, no upstream documentation links leaking into your agent's context.
• CORS enabled, so browser-based agents and web widgets can call it directly.

Errors

Scope and compliance

BusyBee serves public social media data only, and only to users outside mainland China. Do not use it to profile private individuals, scrape private accounts, manipulate engagement, or train models on personal data without a lawful basis. See the data policy.

License

MIT — see LICENSE. This repository is the client-side MCP server; the hosted service is metered separately.

中文速览

BusyBee MCP 把 TikTok、Instagram、YouTube、X、抖音、小红书六个平台的公开数据，压成 10 个业务级工具 + 一套统一字段，同一台服务器支持四种接入方式：

• 密钥：注册后在控制台建 sk- 开头的串
• 计费：按次，每次返回都带本次花费与余额；失败不扣费
• 报错是干净中文/英文，不暴露任何上游信息