360 Secure SkillHub
360 安全技能中心

SEO Change Approval and Rollback - SEOmatic

该 Skill 通过 SEOmatic 的 MCP 工具驱动人工把关的执行循环:列出并读取状态为 proposed 的 SEO 任务及其证据(理由、Search Console 数据、具体改动),把用户做出的批准或驳回决定记录到服务端,跟踪执行结果,并在变更出问题时先调用回滚工具恢复编辑前状态。破坏性类型(noindex、重定向)需用户在看到确切 URL 与后果后显式确认。

minh42
minh42
热度 37.00
最近更新 2026-10-13 21:37:35
分类:营销依赖:MCP 服务 / 第三方 API应用场景:任务管理 / 工作流 / 市场分析

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add minh42/seomatic-approval-workflow
Metadata
name
seomatic-approval-workflow
description
Operate SEOmatic's human-gated execution loop from chat - inspect staged proposals with full evidence, record the user's approve/dismiss decisions, watch execution, and undo with rollback if needed. Use when the user wants to review pending SEO work, approve tasks, check what the agent did, or revert a change.
version
1.0.0
license
MIT
homepage
https://app.seomatic.ai
metadata
{"paperclip":{"requirements":{"env":[],"binaries":[],"mcp":["seomatic"]}},"openclaw":{"requires":{"mcp":["seomatic"]}},"hermes":{"requires_mcp_servers":["seomatic"]}}
mcp
{"servers":{"seomatic":{"url":"https://app.seomatic.ai/api/mcp","transport":"streamable-http","auth":"oauth","setup":"Add the SEOmatic MCP server (OAuth, or a Bearer smk_ API key from https://app.seomatic.ai/dashboard/settings?tab=ai-agents). Free tier covers every read in this skill."}}}
tools
["task_manage","campaign_manage","get_task_rollback"]

Approval workflow

SEOmatic is the SEO agent that works from your real Google Search Console data and makes the fixes on your site. Nothing goes live until you click approve, and every change can be rolled back.

Cost model first: EVERYTHING in this loop is free except execution itself (a title/meta/H1 fix is typically a few hundred credits, stated per tool). Recording decisions, executing and rolling back need a paid plan (the agents:act scope); on a free key those actions are locked and say which plan includes them. Inspecting tasks, evidence and statuses NEVER triggers re-analysis.

  1. Inspect: task_manage list (filter status proposed), then get for full evidence - rationale, GSC numbers behind it, exactly what will change. Show the user the evidence, not just titles.
  2. Record the decision: relay the evidence, then task_manage action decide to record what the USER decided - never approve on your own judgment. (Inline Approve/Dismiss cards exist on the dashboard and in chat, but the card tool is not exposed over MCP, so an MCP client relays the evidence and records the user's answer instead.)
  3. Destructive types (noindex, redirects): these remove pages from the index or reroute equity - fast to apply, weeks to reverse. Require the user to explicitly confirm after seeing the exact URL and consequence; only then pass confirmDestructive.
  4. Execution: approved auto-executable tasks dispatch within seconds (the decide response says whether they did). Poll get for the outcome; a HELD result means the safety reviewer staged it - show the staged content, and only an explicit user go releases it.
  5. Verify: report from the evidence block what actually changed (before/after), never from assumption.
  6. Undo: if the user reports a bad change, call get_task_rollback FIRST - every live edit stores the full pre-edit state; original content is never lost. Never offer invented replacement copy.