引言
「不要给我一堆警告，给我一份计划。」
这是"一天一个开源项目"系列的第 251 篇文章。今天带你了解的项目是 Claude SEO（AgriciDaniel/claude-seo）。
做过网站的人大概都经历过这种场景：跑完一轮 SEO 审计工具，拿到几百条红黄警告，却不知道先改哪条。Claude SEO 的思路是把"SEO 顾问"做成一个 Claude Code 插件：你给它一个网址，它派出一队专家 Agent 并行检查，最后汇总成一份按影响排序、并说明如何验证是否生效的行动计划。
它也是一个很好的"Claude Code 插件工程化"样本：26 个子技能、19 个 Agent、一批带 SSRF 防护的 Python 脚本、一个 PostToolUse Hook，规模不小，结构却很清楚。
你将学到什么
- Claude SEO 解决什么问题，以及它与传统 SEO 工具的分工
- 三条命令完成安装，以及第一次审计怎么跑
- /seo 命令体系与 19 个专家 Agent 的分工
- 一次 /seo audit 内部如何编排并行 Agent
- 插件的目录结构、运行时隔离、SSRF 防护和 Hook 设计
- GEO / Agentic 这两类面向 AI 搜索的新检查项
前置知识
- 了解 Claude Code 的基本用法，知道什么是 Skill、Subagent、Plugin
- 对 SEO 基本概念（sitemap、canonical、Schema.org、Core Web Vitals）有印象即可
- 读 Python 脚本的能力（剖析部分会提到少量实现细节）
项目背景
项目简介
Claude SEO 是一个免费、MIT 协议的 Claude Code 插件。README 的自我定位很直接：在 Claude Code 里给任意网站做完整 SEO 审计，耗时约 10 到 15 分钟，然后交付一份按重要性排序的行动计划。
它覆盖的范围（据 README 与插件清单）：
- 技术 SEO：抓取、索引、sitemap、robots.txt、重定向、页面速度与 Core Web Vitals
- 内容质量：E-E-A-T、薄内容、主题覆盖
- 结构化数据：检测、校验并生成 Schema.org 标记
- AI 搜索（GEO）：Google AI Overviews、AI Mode、ChatGPT、Perplexity、Gemini 引用你页面的可能性
- 本地与电商、国际化 SEO：Google Business Profile、评论、地图包、商品页、hreflang
作者介绍
- 作者：Agrici Daniel（README 称其为 AI Workflow Architect），并提到有 AI Marketing Hub Pro 社区成员参与
- 项目创建时间：2026-02（GitHub API 显示创建于 2026-02-07）
- 同系列项目：Claude Blog（写作）、Claude Banana（配图）、ChatGPT SEO、FLOW（42 条 AI 提示词的 SEO 框架）
项目数据
数据取自 GitHub API，取数日期 2026-10-28：
- ⭐ GitHub Stars: 22,935
- 🍴 Forks: 3,116
- 📦 版本: v2.5.0（2026-10-22 发布）
- 📄 License: MIT
- 🧑💻 主要语言: Python（约 1.56 MB），其余为 Shell、PowerShell、HTML
- 🌐 官网: claude-seo.md
- 🐛 Open Issues: 68
项目从 2 月创建到 10 月已经迭代到 v2.5，最近三个版本（v2.4.1、v2.4.2、v2.5.0）在 9 月底到 10 月下旬密集发布，说明维护相当活跃。README 里的"1077 个测试通过"是徽章自述，我没有复跑。
主要功能
核心作用
把一套资深 SEO 顾问的检查清单固化成可重复执行的 Agent 流程。它不是爬虫的替代品，README 自己也承认 Screaming Frog 这类工具爬得更快更深、大型 SEO 套件有自己的反链数据库。Claude SEO 的差异在于：
- 解释：每条发现说明依据与影响
- 排序：按影响分级为 Critical / High / Medium / Low
- 对话：审计完可以继续用自然语言追问
- 覆盖 AI 搜索：GEO 和 Agent 可用性是一等公民
使用场景
- 自己的网站上线前体检
- 一条 /seo audit https://你的域名，拿到整站健康分和修复清单
- 单页深挖
- /seo page 针对定价页、落地页做逐项分析
- AI 搜索可见性
- /seo geo 评估页面是否容易被 AI 引用，检查 AI 爬虫能否访问
- 改版前后对比
- /seo drift baseline 先存基线，改版后 /seo drift compare 找出回归
- 代理商 / 自由职业者
- 首次沟通时现场跑审计，拿真实分数和 Top 修复项做提案
快速开始
要求：Claude Code 1.0.33 或更新、Python 3.11 到 3.14。在 Claude Code 里执行：
/plugin marketplace add AgriciDaniel/claude-seo
/plugin install claude-seo@agricidaniel-claude-seo
/seo setup
/seo setup 会创建隔离的 Python 运行时和无头 Chromium，不动系统其余部分；/seo doctor 可随时检查状态。
第一次审计：
claude # 启动 Claude Code
/seo audit https://example.com # 整站审计，输出优先级计划
/seo page https://example.com/pricing # 单页深度分析
/seo geo https://example.com # AI 搜索就绪度
/seo schema https://example.com # 检查并生成结构化数据
/seo local https://example.com # 本地 SEO 与 GBP
也支持手动安装（macOS / Linux 用 install.sh，Windows 用 install.ps1）。README 特意建议 Windows 用 git clone 而不是 irm | iex，理由是后者会被 Claude Code 自己的安全护栏判定为供应链风险，先下载再审阅脚本更稳妥。
核心特性
- 26 个子技能
- 22 个核心技能 + 1 个总调度 + 1 个框架集成（FLOW）+ 2 个扩展镜像
- 19 个专家 Agent
- 例如 seo-technical、seo-content、seo-schema、seo-geo、seo-agentic、seo-drift，按任务复杂度分配 Sonnet 或 Opus
- 并行审计
- 完整审计最多 17 个 Agent 同时工作
- 业务类型自适应
- 先判断是 SaaS、本地、电商、内容站还是代理商，再决定调用哪些 Agent
- Agent 可用性检查（Agentic）
- 覆盖 Lighthouse Agentic Browsing、llms.txt、Markdown 输出、WebMCP 等
- 报告落地为文件
- Markdown、JSON，并可用 /seo google report 生成 PDF
- 可选扩展
- SE Ranking、Profound、Bing Webmaster（含 IndexNow）、Unlighthouse、Banana 图像生成等，不装也能用
- 本地运行
- 报告保存在你的机器上；但要注意页面内容仍会发送给 Claude（见后文"局限"）
项目优势
上表中的时间、价格对比来自 README，README 自己也注明是"典型市场价格，仅作示意"。
为什么值得关注？
- 它展示了"把专业流程写成 Skill + Agent"的完整范式
- 对 AI 搜索（GEO / AEO）的覆盖比多数传统工具更靠前
- 工程上有安全意识：SSRF 防护、隔离运行时、不凭空 pip install
项目详细剖析
目录结构与组件
仓库约 25 个顶层条目，核心几块：
docs/ARCHITECTURE.md 里的结构图可以概括为：seo 总调度技能在中心，周围是 seo-audit、seo-technical、seo-content、seo-schema、seo-geo、seo-local 等子技能；agents/ 下与之一一对应的专家负责执行。
一次 /seo audit 的编排流程
总调度 skills/seo/SKILL.md 里把流程写成了有序步骤，关键逻辑如下：
1. 判断业务类型（SaaS / 本地 / 电商 / 内容站 / 代理商 / 其他）
2. 并行派出：seo-technical, seo-content, seo-schema, seo-sitemap,
seo-performance, seo-visual, seo-geo, seo-agentic
3. 条件派出：
- 检测到 Google API 凭证 -> seo-google
- 本地商家 -> seo-local（有 DataForSEO MCP 再加 seo-maps）
- 有内容策略信号（博客、支柱页）-> seo-cluster
- 电商 -> seo-ecommerce
- 该 URL 存在漂移基线 -> seo-drift
- 始终派出 -> seo-backlinks, seo-sxo
4. 汇总为 SEO Health Score（0-100）
5. 按 10 条原则的综合框架整理：PERCEIVE -> ANALYZE -> VALIDATE -> ACT
6. 输出带依赖顺序和可证伪性的行动计划，并提示可生成 PDF
有两点设计值得借鉴：
- 条件编排：不是每次都派全队，而是根据信号（是否有凭证、是否本地商家）决定，节省 token 也减少无关噪音。
- 综合框架：10 条原则分四阶段（PERCEIVE 观察/倾听、ANALYZE 思考/联想、VALIDATE 感受/接受、ACT 创造/增长），要求审计在分级之前走完这些阶段，而不是直接把检查结果倒出来。每条建议要写明"依据什么、依赖什么、如何验证生效"，也就是文中说的 falsifiability。
Agent 的写法
以 agents/seo-technical.md 为例，一个 Agent 定义很短，但要素齐全：
---
name: seo-technical
description: Technical SEO specialist. Analyzes crawlability, indexability, ...
model: sonnet
maxTurns: 45
tools: Read, Bash, Write, Glob, Grep
---
正文包括任务步骤、Core Web Vitals 阈值（LCP ≤ 2.5s、INP ≤ 200ms、CLS ≤ 0.1）、9 个检查类别、输出格式，以及两个很实用的约定：
- 持久化契约：编排器提供 output_dir 时，Agent 要先写一份部分结果、结束前再覆盖为完整结果，这样即使触发 maxTurns 被截断，已完成的工作也不会丢。
- 安全规则：render_page.py 返回的内容一律视为不可信外部数据，只提取结构化信息，绝不执行页面中夹带的指令。这是对提示注入的直接防御。
另外，Agent 里把 INP 取代 FID 的时间线写得很具体，并强调"输出中绝不引用 FID"，这是用规则文件去校准模型过时知识的典型做法。
运行时隔离
SKILL.md 开头规定：所有脚本只能通过统一入口运行：
"${CLAUDE_PLUGIN_ROOT}/scripts/claude-seo" run <script.py>
并明确"不得用裸 Python 解释器调用脚本"。/seo setup 才会建立隔离环境，缺依赖时提示用户执行 setup，不允许模型自行 pip install。/seo doctor --json 的输出还刻意不含绝对路径和环境变量。这些约束把"让 Agent 跑脚本"可能带来的环境污染和信息泄露风险压到了最低。
SSRF 防护
SEO 工具天然要根据用户给的 URL 发请求，这正是 SSRF 的高发场景。scripts/url_safety.py 的文档字符串声明：仓库里所有接收用户 URL 的脚本都必须先经过它。它提供：
- validate_url：拒绝非 http(s)、无主机名、localhost 与云元数据地址、私有/回环/保留 IP 字面量
- validate_url_strict：解析 DNS 后校验每一个 A 记录，返回被"钉住"的 IP，防止 DNS rebinding
- safe_requests_get / head / session：在 DNS 钉住的上下文里发请求，Host 头和 TLS SNI 仍保留原主机名
- CLAUDE_SEO_LOCAL_TARGETS：显式白名单，允许审计本地开发服务器或 Tailscale 内机器，且只对最顶层 URL 生效，重定向目标和子资源不会放行，云元数据地址即使列入白名单也拒绝
这块对想写"能联网的 Agent 工具"的人很有参考价值。
Hook：写入即校验
hooks/hooks.json 只有一个 PostToolUse Hook，匹配 Edit|Write，调用 hooks/validate-schema.py 检查刚写入的文件，超时 30 秒。含义是：当 Claude 帮你生成或修改包含 Schema 标记的文件时，会自动校验结构化数据，而不是等到上线后才发现错误。
GEO 与 Agentic：面向 AI 的两类检查
- /seo geo：评估页面"容易被引用"的程度、AI 爬虫可达性，以及品牌在 Wikipedia、Reddit、YouTube、LinkedIn 的存在感。README 的立场是：Google 称优化 AI 搜索仍属 SEO，所以先保证被索引、够快、够清晰，再谈 GEO。
- /seo agentic：检查 AI Agent 能否顺利使用你的网站，包括 Lighthouse Agentic Browsing 得分、可访问性树、llms.txt、Markdown 输出与 WebMCP。有 audit | fix | lighthouse | refresh 几种子模式。
局限与注意事项
- 没有 Google 凭证时精度有限：README 说明，不接 Google API 时 Core Web Vitals 只是实验室估算，索引状态是推断的。可用 /seo google setup 接入 PageSpeed、CrUX、Search Console、GA4。
- 数据会发送给 Claude：页面内容和分析数据会像其他 Claude Code 工作一样发给 Anthropic 的模型，"本地运行"指报告和密钥保存在你机器上，不等于内容不出本机。
- 成本：5 个需要更多判断的 Agent 用 Opus，其余用 Sonnet，完整审计比单页检查更耗额度，可在 agents/ 里改模型。
- 滚动 / 点击后才加载内容的页面结果可能有噪音。
- 两个版本：公开版 MIT 免费，另有 AI Marketing Hub Pro 付费社区提供私有版本与更早的新功能，这是项目的商业化路径，选用前心里有数即可。
- 文中的效果数据（如官网 10.8K 次 Search Console 点击、移动端 PageSpeed 99 分）均来自 README 自述，未经我核实。
项目地址与资源
官方资源
- 🌟 GitHub: github.com/AgriciDanie…
- 🌐 官网: claude-seo.md
- 📚 文档: 仓库 docs/ 目录（INSTALLATION、COMMANDS、ARCHITECTURE、TROUBLESHOOTING、MCP-INTEGRATION）
- 🐛 Issue Tracker: github.com/AgriciDanie…
- 🎬 演示视频: README 中附有 YouTube 链接
相关资源
- Claude Blog：github.com/AgriciDanie… （基于 SEO 发现写博客）
- Claude Banana：github.com/AgriciDanie… （SEO 配图）
- FLOW：github.com/AgriciDanie… （42 条 AI 提示词的 SEO 框架）
总结与展望
核心要点回顾
- 把顾问流程做成插件：26 个子技能 + 19 个 Agent，把 SEO 检查清单工程化。
- 输出是计划而非警告：按影响排序，并要求每条建议说明依据与验证方式。
- 面向 AI 搜索：GEO 与 Agentic 检查是它区别于传统工具的地方。
- 工程细节扎实：隔离运行时、SSRF 防护、不可信内容规则、写入即校验的 Hook。
- 理性看待：爬取规模和反链数据不及专业工具，页面内容会发送给模型，且存在付费社区的商业分层。
适用人群
- 想快速给自己网站做体检、又不想买订阅的独立开发者与站长
- SEO 代理商和自由职业者，需要首次沟通时现场出审计结果
- 想学习如何写大型 Claude Code 插件（Skill 编排、Agent 约束、脚本安全）的开发者
学习建议
先用 /seo doctor 确认环境，再在自己的站点跑一次 /seo page，对照报告里的依据去验证几条结论。想学工程实现，建议按这个顺序读：skills/seo/SKILL.md（编排）→ agents/seo-technical.md（Agent 写法）→ scripts/url_safety.py（安全设计）→ docs/ARCHITECTURE.md（整体结构）。
关于取数：我读取了 README、plugin.json、总调度 SKILL.md、seo-technical Agent、hooks.json、docs/ARCHITECTURE.md 和 url_safety.py 的开头部分；其余 Agent 与脚本只看了文件清单，细节以仓库为准。
欢迎访问 PrimeSkills —— 一个精心策划的 AI Agent 与技能市场，所有内容均经过真实企业级工作流验证。没有噱头，只有真正有效的东西。
更多实用知识和有趣产品，欢迎访问我的个人主页
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Introduction
"Give me no warning, give me a plan."
This is the 251st article in the "One Open Source Project a Day" series. The project that I'm showing you today is Claude SEO.
Those who have done the website probably have experienced this: after running a round of SEO auditing tools, they get hundreds of red and yellow warnings, but they don't know which one to change first. Claude SEO's idea is to make the "SEO Consultant" a Claude Code plugin: you give it a web site, it sends a team of experts, Agent, to check in parallel, and then it is put together as an impact-ordered action plan and how to validate whether it works.
It's also a good "Claude Code Plug-in Engineering" sample: 26 individual skills, 19 Agents, a collection of Python scripts with SSRF protection, a PostToolse Hook, which is not small, but is well structured.
What will you learn?
- Claude SEO solves what and the division of labour with traditional SEO tools
- Three orders to complete the installation and how to run the first audit.
-/seo command system and division of labour between 19 experts, Agent
- Once /seo audit how to organize it internally, Agent
- Directory structure for plugins, running-time segregation, SSRF protection and Hook design
- GEO / Agency These two new checks for AI search
Pre-knowledge
- understand Claude Code's basic usage, know what Skill, Subagent, Plugin.
- Impression of the basic concepts of SEO (sitemap, Canonical, Schema.org, Core Web Vitals)
- The ability to read Python scripts (a small amount of realization is mentioned in the section of the profile)
Project background
Project Profile
Claude SEO is a free, MIT protocol Claude Code plugin. README 's self-location is straightforward: a complete SEO audit of any website in Claude Code takes about 10 to 15 minutes and delivers an action plan in order of importance.
The range it covers (according to README and plugin lists):
- Technology SEO: Capture, Index, Sitemap, Robots.txt, Redirection, Page Speed and Core Web Vitals
- Content quality: E-E-A-T, thin content, thematic coverage
- Structured data: detect, verify and generate Schema.org tags
- AI Search (GEO): Google AI Overviews, AI Mode, ChatGPT, Perplexity, Gemini
- Local and electricians, internationalization
Introduction by author
- Author: Agrici Daniel (ReADME called AI Workflow Architect) and mentioned the involvement of members of the AI Marketing Hub Pro community
- Project creation time: 2026-02 (GitHub API displayed in 2026-02-07)
- Same series of projects: Claude Blog (writing), Claude Banana (mapping), ChatGPT SEO, FLOW (42 AI prompt framework)
Project data
Data taken from GitHub API, take dates 2026-10-28:
- ⭐ GitHub Stars: 22,935
- 🍴 Forks: 3,116
- 📦 Version: v2.5.0 (2026-10-22 publish)
- 📄 License: MIT
- Main languages: Python (approximately 1.56 MB), the remainder being Shell, PowerShell, HTML
- 官 :claude-seo.md
- 🐛 Open Issues: 68
The project was created from February to October and has been inverted to v2.5, with the last three versions (v2.4.1, v2.4.2, v2.5.0), intensive publish between the end of September and late October, indicating that maintenance is active. "1077 Tests Passed" in README is self-described in the badge, and I didn't run back.
Main functions
Core role
A list of senior SEO consultants is consolidated into a duplicate Agent process. It is not an alternative to reptiles, and README itself recognizes that tools such as Screaming Frog climb faster and deeper, large-scale SEO packages have their own anti-chain databases. The difference between Claude SEO is:
- Explanation: basis and impact of each finding
- Sorting: Critical / High / Media / Low
- Dialogue: the audit can continue in natural language
- Overwrite AI Search: GEO and Agent usability are first class citizens
Use scene
- A pre-screening on your website.
- a /seo audit https://yourdomainname, get the whole station health and repair list
- One page deep.
-/seo page for itemized analysis of pricing pages, drop pages
- AI Search Visibility
- /seo geo assesses whether pages are easily quoted by AI and checks if AI reptiles are accessible
- Re-profiling comparison.
- /seo drift baseline save baseline first, and reset later/ seo drift compare find return
- Agent / freelancer
- On-site audit at first communication, with real scores and Top repairs proposed
Fast start.
Request: Claude Code 1.0.33 or updates, Python 3.11 to 3.14. In Claude Code:
/plugin marketplace add AgriciDaniel/claude-seo
/plugin install claude-seo@agricidaniel-claude-seo
/seo setup
/seo setup creates isolated Python running and headless Chromium, without moving the rest of the system; /seo doctor can check the status at any time.
First audit:
# Start Claude Code
/seo audit https://example.com # global audit, output priority plan
/seo page https://example.com/pricing# deep analysis
/seo geo https://example.com #AI search readiness
/seo schema https://example.com # Check and generate structured data
/seo local https://example.com # Local SEO and GBP
Manual installation (macOS / Linux install.sh, Windows install.ps1). README specifically recommends that Windows use guit code instead of irm | iex, on the grounds that the latter will be considered a supply chain risk by Claude Code ' s own security fence, which will be downloaded before the script is reviewed.
Core characteristics
- 26 skills.
- 22 core skills + 1 master schedule + 1 frame integration (FLOW) + 2 extended mirrors
- 19 experts, Agent.
- e.g. seo-technical, seo-content, seo-schema, seo-geo, seo-angentic, seo-drift, distribution by task complexity Sonnet or Opus
- Parallel audits
- Up to 17 full audits Agent works at the same time
- Business type adaptation
- Let's decide if it's SaaS, local, electric, content station or agent, then which one to call.
- Agent Availability Check
- Overwrite Lighthouse Agency Browning, llms.txt, Markdown output, WebMCP, etc.
- The report landed as a document.
- Markdown, JSON, and generate PDF with /seo google report
- Optional Extension
- SE Ranking, Profound, Bing Webmaster, Unlighthouse, Banana Image Generation, etc.
- Run locally
- The report is on your machine; note that the page will still be sent to Claude.
Project strengths
The time and price comparisons in the above table are derived from README, which itself indicates a "typical market price, indicative only".
Why should I care?
- It shows the whole paradigm of writing professional processes into Skill + Agent.
- Yes, AI search (GEO / AEO) is ahead of most traditional tools.
- There's safety awareness in the project: SSRF protection, isolation, pip install
Detailed Project Profile
Directory structure and components
The warehouse has about 25 top-level entries, core blocks:
The structure chart in docs/ARCHITECTURE.md can be summarized as follows: seo master movement skills are at the centre, surrounded by seo-audit, seo-technical, seo-content, seo-schema, seo-geo, seo-local, etc., etc., which are implemented by the corresponding expert under agents/.
One-time /seo audit layout process
The process has been written into an orderly process in the General Schedule Skills/seo/SKILL.md, with the following key logic:
1. Determination of type of business (SaaS / local / electrician / content station / agent / other)
2. Parallel missions: seo-technical, seo-content, seo-schema, seo-sitemap,
seo-performance, seo-visual, seo-geo, seo-agentic
3. Conditions:
- Google API vouchers detected - > seo-google
- Local - > seo-local (with DataForSEO MCP plus seo-maps)
- Content strategy signals (blogs, support pages) - > seo-cluster
-The electrician--
- This URL has a drift baseline - >seo-drift
-Seo-backlinks, seo-sxo
4. Summarized as SEO Health Score (0-100)
Integrated framework organized according to 10 principles: PERCEIVE - > ANALIZE - > VALIDATE - > ACT
6. An action plan for the dependencies and perjury of the output belt, with an indication that PDF can be generated
Two things are worth drawing on:
- Conditionality: not every time a full team is sent, but depending on the signal (whether there is a certificate, whether there is a local business) the savings token also reduce the noise.
- Integrated Framework: 10 Principles have four phases (PERCEIVE observation/hearing, ANALIZE reflection/linking, VALIDATE perception/acceptance, ACT creation/growth), requiring audits to complete these stages before grading, rather than to reverse the results directly. Each recommendation is to state "what is the basis, what is relied upon, and how it is validated to be effective", that is, what the text says.
Agent
In the case of agents/seo-technical.md, an Agent is defined as a short but complete:
---
name: seo-technical
description: Technical SEO specialist. Analyzes crawlability, indexability, ...
model: sonnet
maxTurns: 45
tools: Read, Bash, Write, Glob, Grep
---
The text includes the mission step, the Core Web Vitals threshold (LCP 2.5s, INP 200s, CLS 0.1), 9 check categories, output formats, and two useful engagements:
- The contract of durability: When the organiser provides output dir, Agent will write a partial result and then cover the full result before the end, so that even if the maxTurns are triggered, the work done will not be lost.
- Security rules: anything returned by render page.py is considered to be untrustworthy external data, extract only structured information and never execute the instructions contained in the page. It's a direct defense to the tip.
In addition, the time line used in Agent to replace FID is specific and emphasizes that "FID is never quoted in the output" as a typical way of calibrating outdated knowledge of models with a rule file.
Run-time segregation
SKILL.md begins by stating that all scripts can only be run through a single entrance:
"${CLAUDE_PLUGIN_ROOT}/scripts/claude-seo" run <script.py>
And make it clear that "No nud Python interpreter calls script." /seo setup will create isolation environments, prompt users to execute setup when there is a lack of dependence, and the model is not allowed to pip install. /seo doctor-json's output also deliberately excludes absolute pathways and environmental variables. These constraints minimize the risk of environmental pollution and information leaks that Agent could bring.
SSRF Protection
The SEO tool is naturally requested according to the URL given by the user, which is the high-growth scene of the SSR. Document string for scripts/url safety.py states that all scripts of the recipient URL in the repository must pass first. It provides:
- Validate url: Reject non-http(s), no hostname, localhost and cloud metadata address, private/return/maintenance IP Volume
- Validate url strict: parse the DNS and verify every A record, return to the "cruising" IP to prevent DNS rebinding
-Safe requests get / head /session: request in context nailed by DNS, Host Head and TLS SNI remain the host First Name
- CLAUDE SEO LOCAL TARGETS: Visible white list, allowing the audit of locally developed servers or Tailscale machines to take effect only on top floor URLs, re-directed targets and sub-resources will not be released and cloud metadata addresses will be denied even if they are added to the white list
This piece is very useful for those who want to write "Agent Tool for Networking."
Hook: Writing to Verify
Hooks/books.json has only one PostTooluse Hook, matching Edit|Write, calling Hooks/validate-schema.py to check the files just written, timeout 30 seconds. Meaning: when Claude helps you generate or modify files containing Schema tags, it automatically validates structured data, rather than wait until you get on the line to find the error.
GEO and Agenic: Two types of checks for AI
- /seo geo: Assess the extent to which the page is "easy to be quoted", AI reptiles 'feasibility, and the existence of brands in Wikipedia, Reddit, YouTube, LinkedIn. README's position is that Google says optimising AI is still SEO, so make sure that it is indexed, fast enough, clear enough, and then talk about GEO.
- /seo andentic: Checking if AI Agent can successfully use your website, including Lighthouse Agency Browning, accessable trees, llms.txt, Markdown output and WebMCP. There are several sub-models of audit |fix |lighthouse | refresh.
Limitations and attention
- Limited precision when no Google certificate is available: README indicates that, when Google API is not taken, Core Web Vitals is only a laboratory estimate and index status is extrapolated. Use /seo google setup to access PageSpeed, Crux, Search Console, GA4.
- The data will be sent to Claude: page content and analysis data will be sent to Anthropic models like other Claude Code jobs, where "local run" reports and keys are kept on your machine, not without content.
- Cost: 5 Agents requiring more judgement, using Opus, and the rest using Sonnet, complete auditing is more quota than single-page inspection, and models can be modified in context.
- The result of the scrolling / clicking to load content may be noise.
- Two versions: an open version of MIT free of charge, and the community of AI Marketing Hub Pro Z105Z provides a private version and a new feature for the earlier, which is the commercial path of the project, which can be counted before selection.
- The effects data in the text (e.g., 10.8 K hits from Search Console, move-end PageSpeed 99 points) are from README, without verification by me.
Project address and resources
Official resources
- 🌟 GitHub: github.com/AgriciDanie…
- 官 :claude-seo.md
- Documentation: repository docs/ catalogues (INTALLATION, COMMANDS, ARCHETECTURE, TROUBLESHOOTING, MCP-INTEGRAPION)
- 🐛 Issue Tracker: github.com/AgriciDanie…
- 🎬 Presentation video: README with YouTube links
Related resources
-Claude Blog:github.com/AgriciDanie...
-Claude Banana:github.com/AgriciDanie... (SEO)
- FLOW:github.com/AgriciDanie... (42 AI prompt SEO framework)
Summary and outlook
Review of core elements
- Plug-in of consultant processes: 26 sub-skills + 19 Agents, engineering of SEO checklists.
- Output is a plan rather than a warning: rank by impact and require each recommendation to state the basis and manner of certification.
- For AI search: GEO and Agenic check is where it differs from traditional tools.
- The details of the project are solid: stand-alone running, SSR protection, implausible content rules, write-in or verify, Hook.
- Rational view: the scale of the climb and the counter-chain data are less than professional tools, the page is sent to models and there are commercial layers in the paid community.
Population applicable
- Independent developers and stations who want to quickly examine their websites without buying them. Long
- SEO agents and freelancers, who are required to communicate for the first time, make a live audit.
- Developers who want to learn how to write large Claude Code plugins.
Learning advice
Validate several conclusions against the basis of the report by first confirming the environment and then running at one of its stations. In order to learn how the project works, it is suggested that it be read in this order: skills/seo/SKILL.md (formulation) agents/seo-technical.md (Agent) scripts/url safety.py (security design) docs/ARCHITECTURE.md (corporate structure).
In terms of numbers: I read the beginnings of README, plugin.json, General Schedule SKILL.md, Seo-technicAgent, books.json, docs/ARCHITECTURE.md and url safety.py; the rest of Agent and scripts only look at the list of documents, depending on the warehouse.
Welcome to PrimeSkills - a well-planned AI Agent and skills market, all of which has been validated by real enterprise-level workflow. There's no twitch, there's only something that really works.
More practical knowledge and interesting products. Welcome to my home page.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.