TinyFish：AI Agent 的 Web 基础设施，Search + Fetch 免费，一行装进 Claude Code

Agent 需要上网，但上网很难

给 AI Agent 提供实时 Web 数据是一个比看起来更难的问题：大多数网页是动态渲染的，有反爬机制，返回的 HTML 里 90% 是导航栏、脚本和广告——直接喂给模型，token 暴涨，信息密度极低。

TinyFish 解决这个问题：它是一套专门为 AI Agent 设计的 Web 基础设施，把「搜索」「抓取」「多步骤自动化」「托管浏览器」打包成四个简洁的端点，背后是真实浏览器渲染 + 内置反检测 + 干净的结构化输出。

客户包括 Google Hotels、DoorDash、ClassPass、Amazon。

好消息：Search 和 Fetch 现在永久免费。

---

四个端点，四个场景

端点	做什么	最适合	速度	价格
Search	实时结构化网页搜索，返回 JSON	任何需要检索的场景；Drop-in 替代 RAG 检索层	< 0.5s	免费
Fetch	任意 URL → 干净 Markdown/JSON/HTML	读取特定页面，给模型喂干净内容	几秒	免费
Agent	URL + 自然语言目标 → 结构化 JSON	多步骤流程、复杂任务、结构化数据提取	10s-数分钟	按量计费
Browser	托管云浏览器，接入你的 Playwright/Selenium	深度定制 Agent 和脚本	实时	按量计费

选择逻辑（来自官方文档）：

• 需要搜索结果列表 → Search
• 已有 URL，要读取页面内容 → Fetch
• 需要 Agent 在网站上完成一个工作流 → Agent
• 需要直接控制浏览器跑自定义脚本 → Browser

---

安装方式一：Agent Skill（Claude Code / Codex / Cursor）

一行命令，把 TinyFish 能力装进任何支持 Skill 的 AI 编码工具：

npx skills add github.com/tinyfish-io/tinyfish-cookbook --skill use-tinyfish

安装后，Agent 会：

• 自动知道什么时候该用 Search vs Fetch vs Agent
• 无需用户说「用 TinyFish」——只要请求涉及实时 Web 信息，Skill 就会触发
• 通过 CLI 调用，结果写到文件系统而不是消耗模型 context window

Skill 的触发词（不需要显式说）：

• 搜索/发现类：search, find, look up, research, compare, latest, current, news, pricing, docs
• URL/页面类：fetch, read, summarize, extract from this page, inspect this URL
• 来源支撑类：answer using web sources, verify a fact, check if something changed
• 网站操作类：interact with a site, click through, fill forms, collect structured data

在 skills.sh/tinyfish-io/tinyfish-cookbook/use-tinyfish 可以预览完整 Skill 内容。

---

安装方式二：MCP Server

在 Claude Code、Cursor、Codex、ChatGPT Desktop 或任何 MCP-aware 客户端里添加配置：

{
  "mcpServers": {
    "tinyfish": { "url": "https://mcp.tinyfish.ai" }
  }
}

配置后，Claude 可以直接通过 MCP 工具调用 Search 和 Fetch，无需任何额外代码。

---

安装方式三：CLI

npm install -g @tiny-fish/cli
tinyfish auth login

# 搜索
tinyfish search query "latest Claude model benchmarks"

# 抓取页面
tinyfish fetch content get https://anthropic.com/news

# 搜索学术论文
tinyfish search query "4D Gaussian Splatting 2026" --domain research_paper

CLI 把结果写到文件系统而不是 stdout，token 使用效率更高。

---

安装方式四：REST API

先拿 API Key：agent.tinyfish.ai（Search + Fetch 永久免费，无需信用卡）

# Search
curl "https://api.search.tinyfish.ai?query=AI+agent+tools+2026" \
  -H "X-API-Key: $TINYFISH_API_KEY"

# Fetch — 单 URL
curl -X POST https://api.fetch.tinyfish.ai \
  -H "X-API-Key: $TINYFISH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://github.com/trending"]}'

# Agent — 多步骤任务（SSE 流式）
curl -N -X POST https://agent.tinyfish.ai/v1/automation/run-sse \
  -H "X-API-Key: $TINYFISH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://news.ycombinator.com",
    "goal": "Find the top 5 AI-related stories today. Return JSON with title, URL, and points."
  }'

---

安装方式五：SDK

# Python
pip install tinyfish

# TypeScript
npm install @tiny-fish/sdk

from tinyfish import TinyFish

client = TinyFish(api_key="YOUR_API_KEY")

# Search
results = client.search("best AI coding tools 2026")
for r in results:
    print(r.title, r.url)

# Fetch
pages = client.fetch(["https://anthropic.com/news"])
print(pages[0].markdown)

两个 SDK 完整覆盖 Search、Fetch、Browser、Agent 和 Vault（Agent 级别的认证凭证和会话内存）。

---

Search API 的高级参数

Search 不只是简单检索，有几个实用的进阶参数：

# 地理定向（中文结果，针对中国市场）
curl "https://api.search.tinyfish.ai?query=AI工具&location=CN&language=zh-CN" \
  -H "X-API-Key: $TINYFISH_API_KEY"

# 新鲜度过滤（最近 24 小时）
curl "https://api.search.tinyfish.ai?query=claude+model+update&recency_minutes=1440" \
  -H "X-API-Key: $TINYFISH_API_KEY"

# 学术论文搜索
curl "https://api.search.tinyfish.ai?query=4DGS+reconstruction&domain_type=research_paper&pub_year_min=2025" \
  -H "X-API-Key: $TINYFISH_API_KEY"

# 新闻搜索
curl "https://api.search.tinyfish.ai?query=Anthropic+Claude&domain_type=news&after_date=2026-08-01" \
  -H "X-API-Key: $TINYFISH_API_KEY"

# 附加搜索意图（帮助系统更准确理解需求）
curl "https://api.search.tinyfish.ai?query=tinyfish+SDK&purpose=Find+installation+guide+for+Python" \
  -H "X-API-Key: $TINYFISH_API_KEY"

---

实用场景

竞品价格监控：Fetch 定期抓取竞品定价页 → 结构化 JSON → 送进数据库

研究 Pipeline：Search 检索 ArXiv 论文 → Fetch 获取摘要全文 → 模型总结

GitHub 趋势追踪：每天跑 Fetch 抓 GitHub Trending → 对比昨日结果，提取新上榜项目

多步骤表单填写：Agent 端点处理需要登录和点击的工作流，直接返回结果 JSON

Cookbook 里的现成 demo：

• lego-hunter — 跨 15+ 零售商追踪稀有乐高库存
• silicon-signal — 半导体供应链 + 交货期信号
• research-sentry — 语音优先的学术研究助手，扫描 ArXiv 和 PubMed
• tinyskills — 从文档、GitHub 和博客生成 SKILL.md（正是 TinyFish 的自指 demo）

---

总结

TinyFish 做了一件对 AI Agent 开发者很实际的事：把访问 Web 从「需要自己搭」变成「API 调用」，内置真实浏览器渲染、反检测、token 效率优化。Search 和 Fetch 永久免费，意味着大多数使用场景（检索 + 读页面）零成本。

对于 Claude Code 用户，一行 npx skills add 就能让 Agent 在需要时自动拿到实时 Web 数据，不需要手动触发，不需要写额外代码。

官网: tinyfish.io
文档: docs.tinyfish.ai
Cookbook: github.com/tinyfish-io/tinyfish-cookbook
Skill: skills.sh/tinyfish-io/tinyfish-cookbook/use-tinyfish
Discord: discord.gg/tinyfish

AI Agents Need the Web — But the Web Is Hard

Giving AI agents real-time web data is harder than it looks: most pages are dynamically rendered, have anti-bot measures, and return HTML where 90% is nav bars, scripts, and ads — feed that directly to a model and tokens spike while information density collapses.

TinyFish solves this: web infrastructure purpose-built for AI agents, packaging “search,” “fetch,” “multi-step automation,” and “managed browser” into four clean endpoints, backed by real browser rendering, built-in stealth, and structured clean output.

Customers include Google Hotels, DoorDash, ClassPass, and Amazon.

Key news: Search and Fetch are now permanently free.

---

Four Endpoints, Four Scenarios

Endpoint	Does	Best for	Speed	Price
Search	Real-time structured web search → JSON	Any retrieval task; drop-in for RAG retrieval	< 0.5s	Free
Fetch	Any URL → clean Markdown/JSON/HTML	Reading specific pages, token-efficient LLM input	Seconds	Free
Agent	URL + natural language goal → structured JSON	Multi-step flows, complex tasks, data extraction	10s–minutes	Metered
Browser	Managed cloud browser for your Playwright/Selenium	Deep-custom agents and scripts	Real-time	Metered

Decision logic:

• Need a list of search results → Search
• Have the URL, want to read the page → Fetch
• Need an agent to complete a workflow on a site → Agent
• Need direct browser control for custom scripts → Browser

---

Install Option 1: Agent Skill (Claude Code / Codex / Cursor)

One command, TinyFish capabilities installed into any Skill-aware AI coding tool:

npx skills add github.com/tinyfish-io/tinyfish-cookbook --skill use-tinyfish

After install, the agent will:

• Know when to use Search vs Fetch vs Agent automatically
• Trigger without the user saying “use TinyFish” — any request involving live web info activates it
• Write results to the filesystem instead of consuming model context window tokens

Skill auto-triggers on: search/find/research/compare/latest/news/pricing, fetch/read/summarize/extract from URL, answer using web sources, interact with a site.

Browse the full Skill at skills.sh/tinyfish-io/tinyfish-cookbook/use-tinyfish.

---

Install Option 2: MCP Server

Add to Claude Code, Cursor, Codex, ChatGPT Desktop, or any MCP-aware client:

{
  "mcpServers": {
    "tinyfish": { "url": "https://mcp.tinyfish.ai" }
  }
}

Claude can then call Search and Fetch directly as MCP tools, no extra code needed.

---

Install Option 3: CLI

npm install -g @tiny-fish/cli
tinyfish auth login

tinyfish search query "latest Claude benchmarks"
tinyfish fetch content get https://anthropic.com/news
tinyfish search query "4D Gaussian Splatting" --domain research_paper

CLI writes results to the filesystem rather than stdout — better for token efficiency.

---

Install Option 4: REST API

Get a free API key at agent.tinyfish.ai (no credit card for Search + Fetch):

# Search
curl "https://api.search.tinyfish.ai?query=AI+agent+tools+2026" \
  -H "X-API-Key: $TINYFISH_API_KEY"

# Fetch
curl -X POST https://api.fetch.tinyfish.ai \
  -H "X-API-Key: $TINYFISH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://github.com/trending"]}'

# Agent (streaming SSE)
curl -N -X POST https://agent.tinyfish.ai/v1/automation/run-sse \
  -H "X-API-Key: $TINYFISH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://news.ycombinator.com",
    "goal": "Find top 5 AI stories today. Return JSON with title, URL, points."
  }'

---

Install Option 5: SDK

from tinyfish import TinyFish

client = TinyFish(api_key="YOUR_API_KEY")

results = client.search("best AI coding tools 2026")
for r in results:
    print(r.title, r.url)

pages = client.fetch(["https://anthropic.com/news"])
print(pages[0].markdown)

Both Python and TypeScript SDKs cover all four endpoints plus Vault (agent-grade credential storage and encrypted session reuse).

---

Advanced Search Parameters

# Geo-targeted (Chinese results for China market)
?query=AI工具&location=CN&language=zh-CN

# Freshness filter (last 24 hours)
?query=claude+update&recency_minutes=1440

# Academic papers
?query=4DGS+reconstruction&domain_type=research_paper&pub_year_min=2025

# News with date range
?query=Anthropic+Claude&domain_type=news&after_date=2026-08-01

# Search intent (helps system understand the goal behind the query)
?query=tinyfish+SDK&purpose=Find+Python+installation+guide

---

Cookbook Demos Worth Running

• tinyskills — generates a SKILL.md from docs, GitHub, and blogs — TinyFish’s self-referential demo
• silicon-signal — semiconductor supply chain + lead-time signals
• research-sentry — voice-first academic research assistant scanning ArXiv and PubMed
• competitor-analysis — live competitive pricing intelligence dashboard

---

Summary

TinyFish does one practically useful thing for AI agent developers: turns accessing the web from “build it yourself” into an API call, with real browser rendering, stealth, and token efficiency built in. Search and Fetch are permanently free, which covers the majority of use cases (retrieval + page reading) at zero cost.

For Claude Code users, one npx skills add line gives the agent automatic access to live web data whenever it needs it — no explicit trigger required, no extra code.

Website: tinyfish.io
Docs: docs.tinyfish.ai
Cookbook: github.com/tinyfish-io/tinyfish-cookbook
Skill: skills.sh/tinyfish-io/tinyfish-cookbook/use-tinyfish
Discord: discord.gg/tinyfish