# 用 3 个免费工具让 AI 写技术周报：GitHub + Hacker News + arXiv（附真实调用）

> 让 Claude 或 ChatGPT 自动收集本周 GitHub 热门项目、Hacker News 首页和 arXiv 最新论文，整理成一份技术周报。本文每一步都是真实调用记录，附提问模板、返回结果和一个常见误区：按相关度搜索拿到的不是「最新」。全程免费。

Published 2026-11-15 · https://toolsocket.top/blog/ai-tech-weekly-free-tools

每周花一两个小时刷 GitHub、Hacker News 和 arXiv，是很多工程师和研究者的固定功课。这件事很适合交给 AI：它能同时查三个来源，再按你关心的方向筛选和总结。

前提是 AI 得能**实时**访问这些网站。这篇教程用 ToolSocket 的三个免费工具来做，每一步都是 **2026 年 11 月 15 日的真实调用**，展开可以看到原始返回。

## 准备

在 Claude、ChatGPT 或 Cursor 里添加 MCP 地址 `https://toolsocket.top/mcp`（[配置方法](https://toolsocket.top/blog/mcp-setup-claude-code-cursor)）。这三个工具都免费，不需要充值。

## 第 1 步：GitHub 上最热的 MCP 项目

> 对你的智能体说: "在 GitHub 上搜「mcp server」，按 star 数列出前 8 个项目，写清楚每个是做什么的。"

Real run of `github#search_repos` with `{"query":"mcp server","sort":"stars","limit":8}` — 1.0 s, free, recorded 2026-11-15:

```json
[
  {
    "full_name": "punkpeye/awesome-mcp-servers",
    "description": "A collection of MCP servers.",
    "url": "https://github.com/punkpeye/awesome-mcp-servers",
    "stars": 97718,
    "forks": 20001,
    "open_issues": 5122,
    "language": null,
    "license": "MIT",
    "topics": [
      "ai",
      "mcp"
    ],
    "default_branch": "main",
    "pushed_at": "2026-11-06T17:59:03Z"
  },
  {
    "full_name": "headroomlabs-ai/headroom",
    "description": "Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers.…",
    "url": "https://github.com/headroomlabs-ai/headroom",
    "stars": 79684,
    "forks": 6238,
    "open_issues": 305,
    "language": "Python",
    "license": "Apache-2.0",
    "topics": [
      "agent",
      "ai",
      "… 18 more"
    ],
    "default_branch": "main",
    "pushed_at": "2026-11-15T07:07:58Z"
  },
  {
    "full_name": "zylon-ai/private-gpt",
    "description": "Complete API layer for private AI applications on local models: RAG, skills, tools, MCP, text-to-sql, and more. Works with any OpenAI-compatible inference se…",
    "url": "https://github.com/zylon-ai/private-gpt",
    "stars": 57636,
    "forks": 7626,
    "open_issues": 14,
    "language": "Python",
    "license": "Apache-2.0",
    "topics": [
      "ai",
      "ai-tools",
      "… 1 more"
    ],
    "default_branch": "main",
    "pushed_at": "2026-11-13T17:13:10Z"
  },
  "… 5 more"
]
```

| 项目 | Star | 语言 | 是做什么的 |
|---|---|---|---|
| punkpeye/awesome-mcp-servers | 97,718 | — | MCP 服务器合集 |
| headroomlabs-ai/headroom | 79,684 | Python | 在工具输出、日志、RAG 片段进入模型前压缩它们，减少 token |
| zylon-ai/private-gpt | 57,636 | Python | 基于本地模型的私有 AI 应用 API 层 |
| microsoft/playwright-mcp | 39,657 | TypeScript | 微软官方的 Playwright 浏览器自动化 MCP |
| DayuanJiang/next-ai-draw-io | 36,953 | TypeScript | 用对话画和编辑 draw.io 图表 |
| github/github-mcp-server | 34,472 | Go | GitHub 官方 MCP 服务器 |
| t8y2/dbx | 30,362 | Rust | 轻量跨平台数据库客户端 |
| PrefectHQ/fastmcp | 28,472 | Python | 用 Python 快速构建 MCP 服务器和客户端 |

按 star 排序能看到「最有名的」，但看不出「本周新冒出来的」。想看新项目，可以让 AI 在关键词里加上 `created:>2026-11-01` 这类 GitHub 搜索语法。

## 第 2 步：本周 Hacker News 首页

> 对你的智能体说: "Hacker News 首页现在前 10 条是什么？按分数排序，一句话说明每条讲了什么。"

Real run of `hackernews#front_page` with `{"limit":10}` — 1.0 s, free, recorded 2026-11-15:

```json
[
  {
    "title": "SpaceX catches the Starship to end Flight 15",
    "url": "https://www.spacex.com/launches/starship-flight-15",
    "hn_url": "https://news.ycombinator.com/item?id=50500963",
    "points": 1069,
    "comments": 712,
    "author": "ls612",
    "created_at": "2026-11-14T21:22:24Z"
  },
  {
    "title": "Lovable apps expose passport details and more than 200K private messages",
    "url": "https://www.crissyfield.de/research/lovable-en/",
    "hn_url": "https://news.ycombinator.com/item?id=50499017",
    "points": 765,
    "comments": 303,
    "author": "armcat",
    "created_at": "2026-11-14T17:43:10Z"
  },
  {
    "title": "Why get an expensive Pi when I have an old phone?",
    "url": "https://so.long.thanks.fish/why-pi-when-i-have-an-old-phone/",
    "hn_url": "https://news.ycombinator.com/item?id=50440709",
    "points": 545,
    "comments": 298,
    "author": "casi",
    "created_at": "2026-11-10T14:16:05Z"
  },
  "… 7 more"
]
```

| 分数 | 评论 | 标题 |
|---|---|---|
| 1,069 | 712 | SpaceX catches the Starship to end Flight 15 |
| 765 | 303 | Lovable apps expose passport details and more than 200K private messages |
| 545 | 298 | Why get an expensive Pi when I have an old phone? |
| 439 | 96 | KDE: The documentary [video] |
| 278 | 137 | A family tree of software job titles |
| 231 | 99 | PyPI's redesigned project pages with clearer security signals are now live |
| 227 | 94 | Project Zero will disclose vulnerabilities 90 days after reporting |
| 219 | 21 | Show HN: DeepSeek V3 in ~200 lines of pure Jax |
| 201 | 104 | Generative AI uses more energy on non-English text |
| 193 | 53 | We're Asking an Appeals Court to Require a Warrant for ALPR Searches |

### 一个常见误区：「搜索」不等于「最新」

如果你问的是「Hacker News 上关于 MCP 的讨论」，AI 会用搜索工具。注意看返回结果的日期：

Real run of `hackernews#search` with `{"query":"MCP","tags":"story","limit":8}` — 1.0 s, free, recorded 2026-11-15:

```json
[
  {
    "title": "Supabase MCP can leak your entire SQL database",
    "url": "https://www.generalanalysis.com/blog/supabase-mcp-blog",
    "hn_url": "https://news.ycombinator.com/item?id=44502318",
    "points": 848,
    "comments": 470,
    "author": "rexpository",
    "created_at": "2025-07-08T17:46:55Z"
  },
  {
    "title": "MCP: An (Accidentally) Universal Plugin System",
    "url": "https://worksonmymachine.substack.com/p/mcp-an-accidentally-universal-plugin",
    "hn_url": "https://news.ycombinator.com/item?id=44404905",
    "points": 808,
    "comments": 350,
    "author": "Stwerner",
    "created_at": "2025-06-28T14:22:10Z"
  },
  {
    "title": "OpenAI adds MCP support to Agents SDK",
    "url": "https://openai.github.io/openai-agents-python/mcp/",
    "hn_url": "https://news.ycombinator.com/item?id=43485566",
    "points": 807,
    "comments": 267,
    "author": "gronky_",
    "created_at": "2025-03-26T18:55:29Z"
  },
  "… 5 more"
]
```

前 8 条大多是 2025 年的帖子。原因是这个搜索**默认按相关度和热度排序，范围是全部历史**，所以最高赞的老帖子排在前面。这对「大家对 MCP 的主要看法是什么」这类问题正合适，但如果你要的是「本周」的讨论，应该：

- 用首页工具 `hackernews#front_page`（上一步），或者
- 提问时明确说「只要最近 7 天的」，让 AI 拿到结果后按日期过滤。

## 第 3 步：arXiv 最新论文

> 对你的智能体说: "在 arXiv 上找最新的 5 篇关于「LLM 智能体使用工具」的论文，用中文各写两句话的摘要。"

Real run of `arxiv#search` with `{"query":"tool use LLM agents","sort":"submittedDate","limit":5}` — 0.8 s, free, recorded 2026-11-15:

```json
[
  {
    "title": "Auditing AI peer reviewers: a benchmark on real scientific papers",
    "authors": [
      "Íñigo Zubeldia",
      "Boris Bolliet",
      "… 3 more"
    ],
    "published": "2026-11-12T16:33:51Z",
    "abstract": "LLM-based peer review is now used at production scale, but its evaluation remains limited. We introduce a benchmark that injects objective, localisable error…",
    "url": "http://arxiv.org/abs/2611.17052v1",
    "pdf": "https://arxiv.org/pdf/2611.17052v1"
  },
  {
    "title": "Skill Adapter: Personalized Modification and Online Routing of General-Purpose Skills",
    "authors": [
      "Qijia Zhuang",
      "Yanyi Gou",
      "… 2 more"
    ],
    "published": "2026-11-12T16:05:23Z",
    "abstract": "Large language model (LLM) agents use external skills to reuse task procedures and experience with tools. General-purpose skills describe shared execution pr…",
    "url": "http://arxiv.org/abs/2611.17007v1",
    "pdf": "https://arxiv.org/pdf/2611.17007v1"
  },
  {
    "title": "Lost in Delegation: Studying the Effectiveness of Task Orchestration in Generative Search",
    "authors": [
      "Tamara Stojanovska",
      "Elisabeth Kirsten",
      "… 2 more"
    ],
    "published": "2026-11-12T09:25:33Z",
    "abstract": "LLMs have quickly evolved from models used in isolation to agentic systems that delegate tasks to external tools and other LLMs to solve complex assignments.…",
    "url": "http://arxiv.org/abs/2611.16526v1",
    "pdf": "https://arxiv.org/pdf/2611.16526v1"
  },
  "… 2 more"
]
```

这次搜索用的是 `sort: submittedDate`（按提交时间），所以拿到的都是 11 月 12 日刚提交的论文：

| 论文 | 一句话 |
|---|---|
| Auditing AI peer reviewers | 在真实论文里注入四类错误，评测 AI 审稿人能发现多少 |
| Skill Adapter | 让通用的智能体技能按用户习惯调整，并在线选择用哪个技能 |
| Lost in Delegation | 研究在生成式搜索里，智能体什么时候把任务委派给工具或其他模型才真正有用 |
| What Can a VLA Learn from the Agent That Calls It? | 把调用方智能体的成功经验蒸馏进视觉-语言-动作模型，导航时不再需要 LLM 规划器 |
| JudgeMix | 用评判模型引导混合轨迹，合成更贴近目标模型的智能体训练数据 |

按时间排序拿到的是「最新」，但不一定「最相关」：比如第一篇讲的是 AI 审稿，和工具使用关系不大。实际使用时可以让 AI 多取 20 篇，再自己筛出最相关的 5 篇。

## 合成一句话，每周跑一次

> 对你的智能体说: "帮我写本周技术周报：① GitHub 上 star 最多的 8 个 MCP 相关项目；② Hacker News 首页前 10 条，标出和 AI 相关的；③ arXiv 最近提交的 20 篇 LLM 智能体论文中，挑最相关的 5 篇写中文摘要。每部分用表格，最后写 3 条本周值得关注的趋势。"

本文的 4 次调用合计耗时约 3.7 秒，**费用 $0**。

如果你用 Claude Code 或 Codex，可以让它把这段流程写成脚本，配合定时任务每周一早上自动生成。

## 常见问题

### 这些工具真的免费吗？

GitHub、Hacker News、arXiv，以及天气、维基百科、汇率、npm、PyPI、图书、词典和 DeFi 数据都免费。通过 REST 调用时，免费工具连 API Key 都不需要。

### GitHub 搜索有频率限制吗？

GitHub 的公开搜索接口有频率限制。正常使用（每分钟几次）不会碰到；短时间内大量调用可能会被 GitHub 暂时限流，稍等一会儿再试即可。

### 能加上中文社区吗，比如知乎、B 站、微博？

能，ToolSocket 也有知乎热榜、B 站搜索和热门、微博热搜等工具，不过是付费的，大多 $0.0015/次/次。在提问里加一句「再看看知乎热榜上有没有相关话题」即可。
