Claude Code for SEO: from first prompt to a team of agents

Claude Code for SEO means using an AI agent that reads your site's files, edits them, runs commands and, through an MCP connector, reads your Search Console, Analytics and Google Business Profile data. Start small: one folder under git, a rule that you read every diff, and read-only data access. Once that loop is reliable, add parallel agents and then several sessions with a lead.

Some people call this "vibe coding SEO" or "vibe SEO": describe the outcome in plain words and let the AI write the code. It is an informal label, not an Anthropic or Google term, and it is only safe with the checks below. Related labels include "Claude SEO", "Claude Code SEO" and "MCP SEO". To start now, connect the ZeusSEO connector, which is included on every plan: the Google Business Profile MCP guide has the setup, and the MCP section below adds the Claude Code commands. If you would rather read about a ready-made expert than build a workflow, there is also Athena, the ZeusSEO AI SEO expert (in early access).

What Claude Code is, and what it can do for SEO

Anthropic describes Claude Code as an agentic coding tool that reads your codebase, edits files, runs commands and integrates with your development tools. It runs in a terminal, an IDE, a desktop app and a browser, and most surfaces need a Claude subscription or an Anthropic Console account.

For SEO, "codebase" means any folder of text: templates, markdown posts, schema snippets, robots.txt, redirect maps and CSV exports. "SEO with Claude" often means chatting in a window; Claude Code works inside a folder on your machine, so it can open files, change them and run checks. It has no view of Google until you add one, as exports or through a connector.

Recommended and suggested tools

Install from the official overview, as a terminal tool, the VS Code extension or the desktop app. Every surface uses the same engine, so your CLAUDE.md files, settings and MCP servers carry over.

A starter CLAUDE.md for a client folder might say:

• Which folders hold templates, content and exports.
• Never edit robots.txt or the redirect map without asking.
• Never invent facts, prices, reviews or statistics.
• Propose Google Business Profile changes; never publish them.

Git and local folders: your undo button

Give each client one folder for site files, exports and reports, and start Claude Code there. In Manual mode it asks before reading or writing outside the folder where it started, as the security page explains.

1. Use git, even for exports. No repo for a CMS site? Keep drafts and CSVs in a folder and run git init.
2. One branch per task, and commit before every run.
3. Read git diff before you accept. Claude Code can commit for you, but the review is yours.

Five SEO workflows to start with

Use one repeatable AI SEO workflow for all five Claude Code SEO jobs: scope, pull data, plan, change files on a branch, run a check, read the diff, commit. Anthropic's best practices recommend planning before coding and giving Claude a check it can run, so it can iterate until the check passes.

1. Technical audit of a site's files

Ask Claude to read templates, robots.txt and sitemap config in plan mode and list findings by file and line. Google says robots.txt is not a way to keep a page out of Google, so flag any Disallow used for that. Our technical SEO audit guide covers the live side.

2. Writing and checking schema

Google's structured data introduction recommends JSON-LD in most cases and says markup must describe content users can see. Have Claude draft it per template, then validate it yourself.

3. Bulk on-page fixes in a repo

Give Claude a CSV of URL and new title and let it edit the matching files. Test on a handful first, then use /batch, which splits a change across 5 to 30 subagents that each work in their own worktree.

4. Keyword-cluster spreadsheets

Export queries to a CSV and ask Claude to write and run a script that groups them by topic and proposes a page for each group.

5. Report prep

Pull numbers through the connector and ask for a short narrative per client. See our SEO ranking report guide for what a client report should hold.

MCP: connect ZeusSEO to Claude Code

MCP, the Model Context Protocol, is an open standard for connecting AI tools to external data sources, and Claude Code supports remote MCP servers over HTTP. MCP for SEO means your agent reads live client data instead of a pasted export. The ZeusSEO connector has one URL and more than 100 tools.

1. Add the server in a terminal, not inside a session:

2. Check it. claude mcp list shows "Needs authentication" until you sign in, as the MCP quickstart describes.
3. Sign in. Start claude, type /mcp, pick zeuseo and choose Authenticate. Your browser opens to the sign-in step described in the Google Business Profile MCP guide.

Add --scope user to use it in every project. If you added the connector in Claude and use that same account for Claude Code, the MCP docs say Claude connectors load automatically, unless an API key is your login. Our sign-in is a standard OAuth flow with PKCE that accepts Claude Code's local callback address.

Reading your own data costs no Staters, ZeusSEO's prepaid usage credits. A few named tools that call our paid providers or AI, such as keyword research and SERP competitor lookups, spend them, and the assistant is told to warn you first. Its instructions also say to confirm content and targets before any publish, edit, delete or reply, and a read-only seat is refused on writes. A free trial has a fair-use cap of 100 API calls. See Staters and AI credits, pricing, and the Search Console and Analytics MCP guides.

API usage: keys, costs and scripts

You sign in with a Claude subscription or an Anthropic Console account. The costs page says the dollar figure in /usage is a local estimate and that subscribers have usage included in their plan.

For scripts, claude -p runs a prompt without an interactive session and accepts piped input, a pattern from the overview:

In --bare mode Claude Code does not use your subscription login, so set ANTHROPIC_API_KEY, as the programmatic usage page explains. If Claude writes scripts that call a data API, keep the key in an environment variable and add the deny rule Read(./.env). For Google data, the connector skips the application wait of your own API project; see Google API vs MCP.

Permissions and hooks: set the guardrails once

Instructions in a prompt or CLAUDE.md shape what Claude tries, but permission rules are enforced by Claude Code, not by the model. Rules are checked in order: deny, then ask, then allow. The permission modes below set how often Claude asks. With Claude Code 2.1.283 or later, auto mode is the built-in starting mode for interactive terminal and VS Code sessions, so for client work press Shift+Tab to reach Manual mode, or start with claude --permission-mode default.

• Allow the read tools: mcp__zeuseo__gsc_*, mcp__zeuseo__ga4_* and mcp__zeuseo__argus_*. The zeuseo prefix is the server name you chose when you added it.
• Keep write tools out of allow rules, so Manual mode asks every time. In auto mode a classifier reviews the call instead, so add an ask rule in /permissions if you want a prompt regardless.
• Deny secrets: Read(./.env).
• Hard blocks: a PreToolUse hook that exits with code 2 stops edits to protected files. The hooks guide says hooks are deterministic, unlike CLAUDE.md.

AI agents for SEO: subagents in parallel

A subagent runs in its own context window with a custom prompt, specific tools and independent permissions, then returns a summary. Define one as a markdown file in .claude/agents/, such as a schema reviewer limited to Read, Grep and Glob so it cannot edit. Anthropic lists several ways to run agents in parallel, including:

One agent acting as lead

The overview describes a lead agent that coordinates the work, assigns subtasks and merges results. Subagents report only to the conversation that spawned them, so that conversation is your single point of contact.

1. Brief the lead: "Use read-only subagents to audit /product, /locations and /blog in parallel. Report by file and line."
2. Merge and edit: the lead removes duplicates, ranks by severity and is the only one that edits, so workers cannot collide.
3. Review: run a fresh reviewer, such as the bundled /code-review skill, on the diff.

Advanced: several sessions, one lead, you as the last gate

This is a working pattern built from documented features, not a built-in mode. The docs describe a built-in lead only for agent teams, which are experimental and off by default; there, teammate permission prompts surface in the lead session, at a higher token cost.

• Same logic: project CLAUDE.md files and permission settings can be shared through version control.
• Separate files: claude --worktree schema-fixes starts a session on its own branch under .claude/worktrees/; the worktrees docs suggest adding that folder to .gitignore.
• Messages: cross-session messaging lets Claude list your other sessions and send them plain text. It needs Claude Code 2.1.224 or later (2.1.234 on native Windows).

1. Commit the rules and split by files: every session starts the same, and no two sessions share a file.
2. Name sessions with --name or /rename; check them with /list-agents, or mention one with @name.
3. The lead assigns by message; workers commit and message a summary back.
4. The lead reviews each diff, runs the checks and rejects out-of-scope edits. You review its short list.

Safe practice and what Google says about AI content

• Read every diff. Anthropic's security page says you are responsible for reviewing proposed code and commands.
• Least privilege, no secrets: allow reads, ask for writes, and never put keys or passwords in prompts, CLAUDE.md or committed config.

A sensible order to adopt all of this

1. Install Claude Code, open one client folder and run git init if needed.
2. Write a short CLAUDE.md, deny Read(./.env) and switch to Manual mode for client files.
3. Run a read-only audit in plan mode.
4. Connect ZeusSEO and allow only the read tools.
5. Make one bulk change on a branch, with a check.
6. Add a read-only reviewer subagent, then a second session in its own worktree, then a named lead.

Managing many locations? The same approach fits our guides on multiple locations, profile optimization and review replies. Your next step is the Google Business Profile MCP guide, which connects the ZeusSEO connector. If you would rather read about a ready-made expert than build your own, there is also Athena (in early access).