avansaber/seo-monster

---

You don’t open a new tool — you just ask, in plain English, inside Claude (or Cursor / Cline / Codex):

You   “What should I write about next?”

🦖 SEOMonster   Pulls your Search Console and surfaces three topics you’re almost ranking for — positions 8–20 with real demand — ranked by opportunity. Say the word and it drafts the brief.

70 tools across Search Console, GA4, PageSpeed, Cloudflare, AI-citation tracking, keyword discovery, and technical SEO — every one returning the same JSON result envelope, every one driven by your own credentials. No new dashboard to learn; you chat with the assistant you already use.

[!NOTE] Published on PyPI as seo-monster (so the command is uvx seo-monster). The import package is seo_mcp; seo-mcp is a dev/local alias. The package ships zero secrets — every credential is resolved at runtime from your own environment or config file.

• Your data, your machine. User-credential-driven; nothing is sent to any third party.
• Read-first, safe by default. Reads always work; the two routine writes (sitemap submit, indexing request) are on by default; the Cloudflare write tools (cache purge, redirects, settings, managed robots.txt) are gated behind SEO_MCP_ALLOW_DESTRUCTIVE, and the riskier ones also need a per-call confirm token.
• Lean. Standard library + the mcp SDK + the Google client libraries; PageSpeed and Cloudflare ride on urllib, no extra HTTP dependency.

Quickstart

[!TIP] Fastest path: the one-click install buttons above. Or add it to any MCP client by hand:

{
  "mcpServers": {
    "seomonster": { "command": "uvx", "args": ["seo-monster"] }
  }
}

Run the one-time Google sign-in once — uvx --from seo-monster seo-monster auth — then ask your assistant to call system_status to confirm what’s connected. Full per-client setup and credentials are in Install and Auth.

How it works

flowchart LR
  U["You — plain English"] --> H["Claude · Cursor · Cline · Codex"]
  H -- "MCP · stdio" --> S["SEOMonster70 tools"]
  S --> GSC["Search Console · GA4"]
  S --> PSI["PageSpeed · CrUX"]
  S --> CF["Cloudflare · IndexNow"]
  S --> AI["AI engines · SERP(optional)"]
  S -. "your creds · your machine" .-> U

SEOMonster is a stdio MCP server: your AI host launches it, it calls Google / Cloudflare / the AI engines with your credentials, and returns a consistent JSON envelope the assistant reads back to you in plain language.

Requirements

For the .mcpb bundle path (Claude Desktop): just Claude Desktop on macOS or Windows. The bundle declares Python 3.11+ as a runtime; Claude Desktop materializes the environment for you. No prior uv install needed.

For the uvx path (Cursor, Cline, Codex, advanced Claude Desktop): Python 3.11 or newer plus uv (which provides uvx). Find the absolute path to uvx with which uvx; GUI hosts do not read your shell profile, so MCP configs need the full path.

Tools

70 tools, grouped by service. All return the same result envelope (see Result envelope). Call system_status first if unsure what is configured. The server also publishes thirteen named workflow prompts.

Cross-service

• system_status - which services are configured/reachable, the Google auth method and scopes, whether destructive mode is on, the full tool catalog, and the list of registered prompts.

Workflow prompts

The server publishes thirteen named MCP prompts (via prompts/list / prompts/get) that chain the granular tools into common SEO workflows. Hosts that surface prompts (Claude Desktop’s slash menu, Cursor’s command palette, Cline’s prompt picker) advertise them automatically.

Why prompts and not megatools: composability. A failed step inside a megatool poisons the megatool’s envelope and the host loses the ability to retry just the failing leg. Prompts hand the host a recipe; each step’s envelope arrives intact at the LLM.

Install

SEOMonster ships two install paths, both fully local:

• .mcpb bundle for Claude Desktop. One-click install, GUI form for credentials, secret-typed inputs stored in the OS keychain. Recommended for most users.
• uvx for Cursor, Cline, Codex, and Claude Desktop power users who prefer to hand-edit MCP config files.

Both paths run the same Python package (seo_mcp) and expose the same 52-tool surface. The difference is only how the host launches the server and how it collects credentials.

Claude Desktop (recommended): .mcpb bundle

Three short steps. The OAuth consent is run once from a terminal (the GUI flow inside Claude Desktop’s MCP subprocess times out before a real user can finish; see Why pre-flight auth? below).

1. Install the bundle. Download seo-monster-0.2.0.mcpb from GitHub releases (or, when listed, from the Claude Directory) and double-click it. Claude Desktop verifies the bundle, runs uv to materialize the Python environment, and shows a configuration form:

Fill the fields, click Save, then toggle the extension on. Quit Claude Desktop completely (⌘Q on macOS) and reopen.

2. Run the one-time OAuth consent from a terminal. Before using any Google-backed tool, run:

uvx seo-monster auth

A browser opens. Approve the requested scopes. The command writes token.json to the path you configured (default ~/.config/seo-monster/token.json) with 0600 permissions, then exits. This step is the recommended pattern; it sidesteps the timeout that Claude Desktop imposes on every tool call.

3. Start a new chat in Claude Desktop and use the tools. Click the 🔧 tools icon in the input box; you should see 52 SEOMonster tools. Try system_status first to verify everything is configured.

Why pre-flight auth?

The OAuth installed-app flow opens a local browser and waits for the user to finish the consent screen. Inside Claude Desktop, MCP servers are launched as subprocesses whose tool calls have a ~30-60 second timeout. Real users do not complete browser consent that fast, so the originating call times out, and since every Google tool retries the flow until a token exists, every call times out in turn. Running uvx seo-monster auth once from a terminal puts the token on disk; from that point on, Claude Desktop’s MCP server just reads the cached token and silently refreshes it as needed.

uvx for Cursor, Cline, Codex (and Claude Desktop power users)

uvx runs the published PyPI package seo-monster in an ephemeral environment. Add the snippet for your host below, using the absolute path to uvx (find it with which uvx; GUI hosts do not read your shell profile).

Cursor (~/.cursor/mcp.json or project .cursor/mcp.json)

{
  "mcpServers": {
    "seomonster": {
      "command": "/Users/me/.local/bin/uvx",
      "args": ["seo-monster"],
      "env": {
        "SEO_MCP_GOOGLE_OAUTH_CLIENT": "/Users/me/.config/seo-monster/client_secret.json",
        "SEO_MCP_GOOGLE_TOKEN": "/Users/me/.config/seo-monster/token.json",
        "SEO_MCP_GA4_PROPERTY_ID": "properties/123456789",
        "PSI_API_KEY": "AIza...",
        "CF_API_TOKEN": "..."
      }
    }
  }
}

Cline (cline_mcp_settings.json)

{
  "mcpServers": {
    "seomonster": {
      "command": "/Users/me/.local/bin/uvx",
      "args": ["seo-monster"],
      "env": {
        "SEO_MCP_GOOGLE_OAUTH_CLIENT": "/Users/me/.config/seo-monster/client_secret.json",
        "SEO_MCP_GOOGLE_TOKEN": "/Users/me/.config/seo-monster/token.json"
      },
      "alwaysAllow": ["system_status", "gsc_search_analytics", "ga4_run_report", "psi_analyze"]
    }
  }
}

alwaysAllow lists read tools so Cline does not prompt on each call. Leave the cache-purge tools off so they always prompt.

Codex (~/.codex/config.toml)

[mcp_servers.seomonster]
command = "/Users/me/.local/bin/uvx"
args = ["seo-monster"]

[mcp_servers.seomonster.env]
SEO_MCP_GOOGLE_OAUTH_CLIENT = "/Users/me/.config/seo-monster/client_secret.json"
SEO_MCP_GOOGLE_TOKEN = "/Users/me/.config/seo-monster/token.json"
SEO_MCP_GA4_PROPERTY_ID = "properties/123456789"

Claude Desktop, direct uvx (advanced)

If you prefer to hand-edit claude_desktop_config.json instead of using the .mcpb bundle, the same snippet shape as Cursor above works.

Auth

The four services authenticate independently. Configure only the ones you use; a tool for an unconfigured service returns a clear AUTH_MISSING error rather than failing the server.

Quick setup (recommended): seo-monster setup

Run seo-monster setup once from a terminal. It interactively collects your Cloudflare token, PageSpeed Insights key, IndexNow key, and the default GSC and GA4 properties, validates what it can against the live APIs, and writes them to ~/.config/seo-mcp/config.toml with 0600 permissions. Your MCP host config then needs no secrets in it:

{ "command": "uvx", "args": ["seo-monster"] }

Two things setup does not do, by design:

• Google OAuth still uses the separate one-time browser step. After setup, run seo-monster auth to complete Google consent (see the next section).
• It never overrides environment variables. Anything set in your host’s env block still wins over the config file, so CI and Docker keep using env vars.

setup is re-runnable: existing values are shown as defaults and kept when you leave a field blank. The sections below document the per-service env vars, which are what setup writes for you and what CI pipelines can set directly.

Google (Search Console + Analytics 4) - OAuth, recommended

This is the lower-friction path: no Cloud service account, no per-property email grants.

1. In the Google Cloud Console, create (or pick) a project and enable the APIs you will use:
   • Search Console API
   • Indexing API (for gsc_request_indexing)
   • Google Analytics Data API (for the GA4 tools)
   • PageSpeed Insights API (only if you want a PSI key; see below)
2. Create an OAuth client of type Desktop app and download the client-secrets JSON.
3. Point the server at it and at a writable token path:
   • SEO_MCP_GOOGLE_OAUTH_CLIENT = path to the client-secrets JSON
   • SEO_MCP_GOOGLE_TOKEN = a writable path where the token will be cached
4. One-time: run uvx seo-monster auth from a terminal. A browser opens; approve the scopes. The command writes token.json (0600) and exits.
5. Subsequent runs (server-side) refresh the token silently. The server never opens a browser; if the cached token is missing, tools return AUTH_MISSING pointing back at the auth command.

The signed-in Google account must have access to the Search Console properties and GA4 properties you query.

Token-cache hardening. The cached token is refresh-capable and equivalent to a long-lived credential for the requested scopes. The server writes it with 0600 and its parent directory with 0700. Keep SEO_MCP_GOOGLE_TOKEN under a directory you control (e.g. ~/.config/seo-monster/) and do not put it on a shared filesystem.

Google - service account (advanced, headless)

For fully headless or server deployments where a browser is not available:

1. Create a service account and download its JSON key.
2. Set SEO_MCP_GOOGLE_CREDENTIALS (or the standard GOOGLE_APPLICATION_CREDENTIALS) to the key path.
3. Grant the service-account email access on each property:
   • Search Console: add it as a user on the property.
   • GA4: add it as a Viewer on the property.

If both OAuth and a service account are configured, OAuth is used.

Coverage note. The OAuth installed-app path is exercised in our validation pass and in production-style smoke tests. The service-account path is documented but not independently validated against a live Cloud project. If you hit issues on the SA path, please open an issue.

Scopes (minimal vs full)

The default consent requests the scopes needed for every tool, including the two writes:

If you only want reads, you can consent to a narrower set (webmasters.readonly + analytics.readonly) and simply not call gsc_submit_sitemap / gsc_request_indexing; calling a write tool without its scope returns SCOPE_INSUFFICIENT with remediation, never a crash.

PageSpeed Insights

PSI works without a key in principle, but in practice the anonymous quota is shared across every caller without a key and is frequently exhausted: a single psi_analyze call against the anonymous endpoint often returns RATE_LIMITED. Treat the anonymous mode as a fallback, not the steady state.

To get reliable PSI access:

1. In Cloud Console, enable the PageSpeed Insights API.
2. Create an API key (Credentials > Create credentials > API key). It takes a minute. The key is free.
3. Set PSI_API_KEY (or use the field in the .mcpb configuration form).

The PSI API only accepts the key as a URL query parameter (not a header), so treat PSI keys as low-sensitivity. Scope the key to the PageSpeed Insights API only and attach no other GCP roles.

Cloudflare

Create an API token at dash.cloudflare.com/profile/api-tokens and set CF_API_TOKEN (and optionally CF_ZONE for a default zone). Grant only the permissions you need:

IndexNow

IndexNow notifies Bing, Yandex, Naver, Seznam, and Yep when a URL is created or updated. Google does not participate, so the IndexNow tools complement rather than replace gsc_request_indexing.

One-time setup

1. Generate a key. Any 8-128 character string of letters, digits, or hyphens (a-z, A-Z, 0-9, -) per the IndexNow spec. Common patterns: a 32-char lowercase hex string (e.g. python -c "import secrets; print(secrets.token_hex(16))") or any random alphanumeric of similar length. Treat it like an API key; do not commit it.
2. Configure SEOMonster by setting SEO_MCP_INDEXNOW_KEY to that string (or use the .mcpb configuration form; the field is marked sensitive and lands in the OS keychain).
3. Host the verification file at https:///.txt. The file body MUST be exactly the key string with no trailing newline, no BOM, no extra whitespace, no HTML wrapper. The Content-Type should be text/plain. Confirm with curl -i https:///.txt before moving on; the response body must be byte-identical to the key.
4. (Optional) Set SEO_MCP_INDEXNOW_KEY_LOCATION if the verification file lives at a non-standard URL (some CDNs rewrite /key.txt paths). The default location is https:///.txt derived from the URLs you submit, so you usually do not need this.

Same-host constraint

Every URL submitted in one indexnow_submit or indexnow_bulk_submit call must share the same host as the verification file. Mixed-host batches are rejected by IndexNow with HTTP 422; indexnow_bulk_submit enforces this client-side and returns INVALID_INPUT before any network call when it detects mixed hosts.

If you have multiple hosts, host a verification file per host and either make separate calls per host or override SEO_MCP_INDEXNOW_KEY_LOCATION per call (the tool does not currently expose per-call override; set distinct env values per session).

Common errors

Verify your setup

After configuring, call system_status to see what is detected. Call it with {"probe": true} to make one cheap live request per configured service and confirm the credentials actually work (GSC lists properties, GA4 runs a 1-row report against the default property, Cloudflare lists one zone, PSI pings the endpoint). With probe off (the default) it does a config-only check and makes no network calls.

Destructive mode

Cache purges affect every visitor, so they are off by default. Set SEO_MCP_ALLOW_DESTRUCTIVE=true to enable cf_purge_cache and cf_purge_cache_all. While off, those tools return DESTRUCTIVE_DISABLED and make no network call.

cf_purge_cache_all (purge the whole zone) carries an extra safeguard: it requires a confirm argument equal to the resolved zone hostname. A missing or mismatched confirm returns CONFIRM_REQUIRED and issues no purge.

The two GSC writes (gsc_submit_sitemap, gsc_request_indexing) are not gated; they are routine, low-blast-radius SEO tasks.

Configuration

Resolution is environment-first, with a TOML file fallback. Environment always wins. The config file is normally written for you by seo-monster setup (with 0600 permissions); you can also write it by hand or set the env vars below.

Config file fallback at ~/.config/seo-mcp/config.toml (or SEO_MCP_CONFIG):

[google]
oauth_client = "/Users/me/.config/seo-mcp/client_secret.json"
token        = "/Users/me/.config/seo-mcp/token.json"
# credentials = "/Users/me/.config/seo-mcp/sa.json"   # service-account alternative

[gsc]
default_site = "sc-domain:example.com"
data_state   = "all"

[ga4]
property_id  = "properties/123456789"

[psi]
api_key = "AIza..."

[cloudflare]
api_token = "..."
zone      = "example.com"

[server]
allow_destructive = false

# Optional v0.9 providers (discovery + AI/GEO). The free GSC/GA4/HTTP core
# works without any of these; each tool degrades gracefully when unset.
[dataforseo]
# login    = "..."
# password = "..."

[openpagerank]
# api_key = "..."

[ai_engines]
# perplexity = "..."
# openai     = "..."
# anthropic  = "..."
# gemini     = "..."

[google_ads]
# developer_token = "..."   # volume alt to DataForSEO; needs adwords-scope OAuth
# customer_id     = "..."

Result envelope

Every tool returns the same shape. On success:

{ "ok": true, "data": { /* tool-specific */ }, "error": null }

On failure:

{
  "ok": false,
  "data": null,
  "error": {
    "code": "AUTH_MISSING",
    "service": "gsc",
    "message": "No Google credentials found for Search Console.",
    "remediation": "Configure OAuth ... or a service-account key. See README > Auth.",
    "docs_url": "https://seomonster.avansaber.com#auth",
    "details": null
  }
}

Error codes:

Development

git clone https://github.com/avansaber/seo-monster
cd seo-monster
uv venv && uv pip install -e ".[dev]"
uv run pytest               # offline test suite
uv run seo-monster          # run the server over stdio
uv run seo-monster auth     # one-time OAuth consent (or `uv run seo-mcp auth`)

The package exposes two console-script aliases: seo-monster (canonical, matches the PyPI distribution) and seo-mcp (a v0.1.x dev alias kept for back-compat). Both invoke the same entry point. As of v0.2.0, invoking the server via seo-mcp emits a one-line stderr deprecation notice; nothing on stdout, so the MCP protocol channel is unaffected. Production configs should use seo-monster; the alias will be removed in a future major release.

Tests are fully offline: they mock at the client layer, so no network and no credentials are needed to run them.

Server identity note. Some MCP host UIs display the server name as seo-mcp and the version as the mcp SDK version (e.g. 1.27.1). The server-name string is the value we passed to Server("seo-mcp") and is kept stable for back-compat; the version readout is a quirk of the SDK (create_initialization_options() does not propagate the package version). The package’s real version is in pyproject.toml and seo_mcp.__version__.

Changelog

Release-by-release notes, including the validation checks each version’s external testing pass should cover, live in CHANGELOG.md.

Privacy

SEOMonster runs entirely on your machine and talks only to the upstream APIs you configure. The maintainers do not see any of your data, credentials, queries, or tool calls. See PRIVACY.md for the full statement.

License

MIT. See LICENSE.