agent-senses
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@agent-sensesfetch this douyin video's caption, transcript, and stats: https://v.douyin.com/abc123/"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
agent-senses
给 agent 用的感官。 | English
大模型不缺脑子,缺的是"看见"和"进去"的能力。一个中文内容链接,人点开就看,agent 却要 撞反爬、绕签名、下流、转写——十几轮工具调用。
这个项目把那十几轮压成一次函数调用。
douyin_get(url) → 文案 + 作者 + 时长 + 互动数据 + 逐句转写 + 本地文件
ds_search(query) → 标题 + URL + 摘要 + 长 summary + 发布日期 + 权威度
ds_fetch(url) → 网页正文(Markdown)
装
pip install -r requirements.txt
# 用系统 Chrome 驱动,不用下 chromium;确认本机装了 Chromeffmpeg 不装也行,会退到 imageio-ffmpeg 自带那份。
转写模型默认 small,从 HF cache 找;下模型:
HF_ENDPOINT=https://hf-mirror.com huggingface-cli download Systran/faster-whisper-small
# 或用环境变量 SENSES_WHISPER_DIR 直接指定模型目录Related MCP server: mcp-china-ecosystem
用
直接跑(调试)
python server.py --selftest "https://v.douyin.com/xxxx/"当 MCP server 挂进 Claude Code
claude mcp add agent-senses --scope user -- <python路径> "<本目录>/server.py"
# 撤销:claude mcp remove agent-senses搜索 / 抓网页(DeepSeek Search & Fetch API)
不依赖 MCP 会话,脚本、管道、定时任务都能直接调:
python dsweb.py search "番茄短故事 千字万金 分成" --count 5 --pretty
python dsweb.py fetch "https://www.ithome.com/0/101/4328.htm" --max-chars 5000query 支持 site: / -site: / "精确短语" / lang:zh,en / after:2025 等操作符。
key 按 DEEPSEEK_API_KEY → ~/.claude/settings.json → ~/.dsh/.credentials.yaml 的顺序找。
计费:各 5 元 / 1000 次成功请求,2026-11-01 前免费。
当库用
from senses import douyin
item = douyin.fetch("https://v.douyin.com/xxxx/", want_video=False)
print(item.author, item.duration)
print(item.text)架构
senses/ 核心库 —— 能力的本体,跟 MCP 无关
models.py 返回给 agent 的数据契约
douyin.py 抖音接入(解析 → 抓取 → 下载 → 合流 → 转写)
web.py 网页搜索/抓取(DeepSeek Search & Fetch API)
media.py ffmpeg 封装
asr.py 语音转写
server.py MCP 外壳 —— 薄,随时可换 CLI / HTTP
dsweb.py 搜索/抓取的 CLI 入口 —— 给不挂 MCP 的脚本用
tests/ 协议层自测为什么库和外壳分开:MCP 协议还在变。核心能力做成独立库,协议换了只换外壳, 本体重写不了。
加下一个平台
照抄 douyin.py 的形状:
resolve(url)—— 链接归一化,拿到平台内部 id_capture(id)—— 把页面数据抠出来(优先拦接口 JSON,别抠 DOM 文本)_download / merge / transcribe—— 这三步是共用的,直接 import在
server.py注册一个xxx_get工具
返回统一用 Item,agent 看到的形状就不变。
踩过的坑(别重踩)
现象 | 原因 | 解法 |
yt-dlp 报 | 抖音 Argus 风控 | 不走 yt-dlp |
| Chromium App-Bound 加密 | 不走浏览器 cookie |
分享页 | SSR 已不下发详情 | 走 Playwright 跑真页面 |
直连 detail 接口 403 |
| 拦页面自己发的请求 |
分片下载 403 | 缺 Referer | 带 |
| PyAV 版本对不上 | 用 wave+numpy 直喂,绕开解码 |
模型加载报 snapshot 不完整 |
| 加载前验 |
GPU 报 | 缺 cuBLAS | 自动退 CPU int8 |
转写里简繁体混着出 | whisper 中文输出不受控 |
|
兜底抠到的作者不是这条视频的 | 播放页会轮播「下一条」 | 只在 DOM 文案和接口文案一致时才用 DOM 补字段 |
| detail 接口没吐全 | 拿 |
边界(当前版本明确不做)
不碰登录态、不做账号操作、不抓私密内容
只读公开内容
转写是"速记"不是"精校"——口音/语速会造成错字,引用前核对
单平台、单条。批量、去重、缓存都还没有
This server cannot be deployed
Maintenance
Related MCP Connectors
660+ data tools for agents: TikTok, Douyin, Instagram, YouTube, X, weather, GitHub. Pay per call.
Web data for agents: YouTube transcripts, screenshots, Google News, WHOIS, jobs, tech stack, more.
TikTok research for AI agents: real TikTok search data, trends, videos, creators and comments in ChatGPT, Claude, Cursor, Codex or any MCP client. 45 tools. See what people search on TikTok each week (Creator Search Insights search volume and 7-day trends), find content gaps with no videos yet, check if a product is trending for TikTok Shop, analyze competitor creators and posts, read comments and replies, pull video transcripts, search hashtags and sounds, and pick fair giveaway winners. Built for marketers, agencies, TikTok Shop sellers and creators. Sign in with OAuth or an API key; 50 free credits.
Web tools for AI agents: URL to clean Markdown, site crawl, screenshots, PDF/Office to text, OCR.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables AI assistants to crawl and extract data from Chinese social media platforms like Bilibili, Xiaohongshu, and Douyin. Provides search, content detail retrieval, and creator information tools with persistent browser sessions and QR code login support.78-
- AlicenseAqualityDmaintenanceBridges AI agents to Chinese internet services like Baidu, Bilibili, and Chinese web page extraction without requiring API keys.3MIT
- AlicenseAqualityFmaintenanceEnables AI agents to download, transcribe, and inspect video or audio URLs from YouTube, TikTok, X, and 1000+ other sites using server-side yt-dlp, residential proxies, and speech-to-text.912 npmMIT
- AlicenseNot gradedqualityBmaintenanceEnables local-first ingestion of Douyin links, favorites, and public Bilibili videos into Markdown, timestamped transcripts, OCR, visual assets, and deduplicated indexes for downstream RAG workflows.1MIT