How to Connect Claude to an SEO MCP Server (and ChatGPT, Codex, Cursor)
Connect Claude, ChatGPT, Codex, Cursor or OpenClaw to an SEO MCP server over OAuth, with the exact steps per client, what to do when the connection fails, and why you should never paste an API token into a client config.
Connecting Claude to an SEO MCP server takes a URL and a browser sign-in. There is no API key to generate and nothing to paste into a config file. The same procedure works for ChatGPT, Codex, Cursor and OpenClaw, because MCP is one open standard rather than a per-client integration.
- THE WHOLE PROCEDURE IS TWO STEPS: Add the server URL in your client's connector settings, then approve the connection in the browser tab that opens. The tools appear in that client immediately afterwards.
- NEVER PASTE A TOKEN: If a server asks you to put an API key in a config file, that credential sits in plain text on disk and does not expire. Prefer OAuth, which gives a scoped grant you can revoke from either side.
- CLAUDE.AI AND DESKTOP: Customize > Connectors > add a custom connector, then paste the server URL. Claude Code uses claude mcp add --transport http instead.
- CHATGPT NEEDS DEVELOPER MODE: Enable developer mode first, then add a custom connector with the same URL and complete the OAuth prompt. This is the one client with an extra gate.
- MOST FAILURES ARE ONE OF FOUR THINGS: A 401 means the grant is gone, a 403 means a missing scope, a paid-plan error means the account at consent was not the paid one, and no tools after approval means the browser tab never returned to the client.
- CHECK THE PLAN BEFORE YOU START: CrawlRaven includes MCP with every paid plan, from $49 at launch, and free accounts cannot approve a connection. Other vendors gate it well above their advertised entry tier.
If the tools do not appear after approval, the token exchange did not finish. Reconnect and let the browser tab return to the client on its own before you close it.
Connecting Claude to an SEO MCP server is a URL and a browser sign-in, not an integration project. The steps below are the ones CrawlRaven documents, per client, along with the four failures that account for most bad connections. Try CrawlRaven free: 1 site, no credit card →
Before you start
Three things decide whether this works, and all of them are worth checking before you open any settings panel.
- An MCP-capable client. Claude on claude.ai, Claude Desktop and Claude Code, plus ChatGPT, Codex, Cursor and OpenClaw. The protocol is open, so newer clients tend to work without the server changing anything.
- The server URL. For CrawlRaven that is
https://mcp.crawlraven.com/mcp. Every client below takes the same string. - An account that qualifies. CrawlRaven includes MCP with every paid plan and free accounts cannot approve a connection. Other vendors gate it higher than their advertised entry tier, which we broke down in our comparison of SEO MCP servers.
The procedure, in two steps
Every client below is a variation on the same two steps, which is the point of a shared protocol. Learn it once and the client stops mattering.
- Add the server URL in the client's MCP or connector settings. The client fetches the server's authorization metadata from that URL and works out how to sign you in.
- Approve the connection in the browser tab that opens, then let that tab hand control back to the client on its own. The tools appear in the client immediately afterwards.
There is no third step. If a guide tells you to generate an API key and paste it somewhere, it is describing a weaker server, not a further part of this one.
Connect Claude.ai and Claude Desktop
Go to Customize > Connectors, add a custom connector, and paste the server URL:
https://mcp.crawlraven.com/mcpClaude opens a browser sign-in, you approve, and the tools land in the conversation. This is the shortest path of the six and the one to use if you are only trying it out.
Connect Claude Code
Add the server from the terminal:
claude mcp add --transport http crawlraven https://mcp.crawlraven.com/mcpThen authenticate, either from /mcp inside a session or with claude mcp login crawlraven. The browser sign-in is the same one the other clients use.
Connect ChatGPT
ChatGPT is the one client with an extra gate. Enable developer mode first, then add a custom connector with the same server URL and complete the OAuth prompt.
Nothing about the server changes for ChatGPT. It advertises OAuth 2.0 Protected Resource Metadata and Client ID Metadata Documents, which ChatGPT prefers, alongside Dynamic Client Registration for clients that use it instead.
Connect Codex
Add the Streamable HTTP server URL, then sign in from the CLI:
codex mcp login crawlravenConnect Cursor
Add a remote Streamable HTTP MCP server with the CrawlRaven URL and choose OAuth when prompted.
https://mcp.crawlraven.com/mcpConnect OpenClaw
Add the server with OAuth, then log in:
openclaw mcp add crawlraven --url https://mcp.crawlraven.com/mcp \
--transport streamable-http --auth oauth
openclaw mcp login crawlravenVerify the connection actually works
A client that lists the tools has not proved it can call them. The fastest real check is to ask for your websites, because list_websites returns the website_id every other tool needs.
Ask the agent: which sites do I have in CrawlRaven? A list means the grant works, the account is paid, and the scopes are right. An empty result with no error usually means you approved consent against a free account.
If the connection is more trouble than the question is worth, the filters in Search Console answer plenty on their own. Our free GSC regex generator builds the query filters without any of this setup.
Why OAuth matters more than the setup time it saves
The convenience argument for OAuth is obvious and slightly beside the point. The security argument is the one worth internalising, because the two approaches fail very differently.
- ✓OAuth grants are scoped, so a client gets only the permissions it asked for and you approved
- ✓Grants are revocable from either side, and revoking takes effect immediately
- ✓Refresh tokens rotate, so ordinary renewal never interrupts the connection
- ✓Nothing sensitive is written to disk in plain text
- ✓S256 PKCE means an intercepted authorization code is useless on its own
- ✗A pasted API key sits in a config file in plain text, readable by anything on the machine
- ✗Pasted keys usually carry full account scope, not just what the tool needs
- ✗They do not expire when you stop using the client, and they survive being copied into a repo
- ✗Rotating one means editing every config it was pasted into, and remembering where those are
When it does not work
Most failed connections are one of a small set of causes, and the error usually names it. These are CrawlRaven's documented responses; other servers differ in wording but rarely in substance.
What the error codes mean
| Code | What it means | What to do |
|---|---|---|
| 401 | Token missing, invalid, expired, revoked, or issued for a different audience | Let the client refresh, or reconnect if the grant is gone |
| 403 | Valid token without the scope the tool needs, or a host or browser origin the server rejects | Reconnect and approve the missing scope |
| forbidden | The account behind the grant has no paid entitlement, so tools return no data | Restore the plan on that account, or reconnect against a paid one |
| 400 | Malformed authorization request | Start the connection again from the client |
| rate_limited | A tool error carrying a retry hint rather than an HTTP status | Wait for the hinted interval instead of retrying in a loop |
Common symptoms and their causes
- Paid plan required. The signed-in user has no paid account, or the account chosen at consent is not the paid one. Switch accounts, or pick a plan from Billing.
- Connection shows as paused. The account behind an active grant lost its paid entitlement. Restore the plan and the grant resumes without reconnecting.
- Authorization request expired. Authorization transactions are deliberately short-lived. Start again from the client.
- No tools after approval. The client did not finish its redirect and token exchange. Reconnect and let the browser tab return to the client before closing it.
- Client registration failed. The client is outside what the server accepts: CIMD or DCR, authorization code flow, S256 PKCE, and HTTPS or loopback callbacks. Update the client, then retry.
- Wrong account. A different account was selected during consent. Revoke the grant, reconnect, and pick the intended one.
What a connection cannot do
Worth knowing before you connect a client to anything holding client data. With CrawlRaven every tool in the current catalog fetches data, and a connection cannot edit websites, run exports, reach billing or administration, or call arbitrary API endpoints.
Do not generalise that to every server. Some WordPress MCP plugins are write-capable by design, which is exactly why they are useful and exactly why they deserve a narrower scope than a read-only analytics server. Read the tool list before connecting anything to a production site.
One thing a connection does do: your data reaches the client and its model provider, processed under their policies. That is the point of connecting, and it is still worth deciding deliberately which clients you are willing to share Search Console data with.
How to disconnect
Revoke the grant from the CrawlRaven side, from the client, or both. Revocation takes effect immediately, and expired or revoked grants cannot be reactivated, so reconnecting means starting a new connection rather than restoring the old one.
The full tool catalog, scopes, parameters, pagination limits and error reference live in the MCP server documentation. If you have not picked a server yet, start with the nine worth connecting to and what each costs before it works.
Frequently asked questions
How can I connect Claude to an SEO MCP?
Add the server URL in Claude's connector settings, then approve the connection in the browser tab that opens. On claude.ai and Claude Desktop that is Customize > Connectors > add a custom connector, then paste the URL. On Claude Code it is a single command, claude mcp add --transport http crawlraven https://mcp.crawlraven.com/mcp, followed by authenticating from /mcp or with claude mcp login crawlraven. There is no API key to generate and nothing to paste into a config file.
Do I need an API key or token to connect Claude to an SEO MCP server?
Not with a well-built server. CrawlRaven uses OAuth authorization code with S256 PKCE, so your client discovers the authorization metadata from the server URL and opens a browser sign-in. Never paste a personal access token, a Google token, or a browser cookie into a client configuration. A pasted key is a long-lived credential sitting in plain text on disk, and it does not expire when you stop using it.
Does the same procedure work for ChatGPT?
Yes, with one extra step. ChatGPT requires developer mode to be enabled before you can add a custom connector, after which you paste the same server URL and complete the OAuth prompt. MCP is an open standard rather than a per-client integration, so the server does not need to do anything different for ChatGPT than it does for Claude. Codex, Cursor and OpenClaw follow the same pattern.
Why do no tools appear after I approved the connection?
The client did not finish its redirect and token exchange, which usually means the browser tab was closed before it returned to the client. Reconnect and let the tab close itself. If the tools still do not appear, check whether the account you selected at the consent screen is the one with the paid plan, because a grant against a free account connects successfully and then returns no data.
What does a 401 or 403 mean on an MCP connection?
A 401 means the token is missing, invalid, expired, revoked, or was issued for a different audience, so let the client refresh or reconnect if the grant is gone. A 403 means the token is valid but lacks the scope the tool needs, or the server rejected the host or browser origin, so reconnect and approve the missing scope. Neither is usually a server fault.
Can a connected AI agent change anything in my account?
With CrawlRaven, no. All eight tools fetch data, and a connection cannot edit websites, run exports, reach billing or administration, or call arbitrary API endpoints. Write tools are a later release with their own consent step. Do not assume this of every server: some WordPress MCP plugins are write-capable by design, so read the tool list before you connect anything to a production site.
What happens to my connection if my plan lapses?
With CrawlRaven the connection pauses rather than disappears. The client stays listed and you can still revoke it, but tools stop returning data until the account has a paid entitlement again. Restoring the plan resumes the grant without reconnecting. Free accounts cannot approve a connection in the first place.
Which SEO MCP server should I connect Claude to?
It depends on whose data you want. For your own Search Console, GA4 and crawl data, CrawlRaven is the closest fit and is included with every paid plan. For third-party keyword and backlink research, SE Ranking, Ahrefs, Semrush and DataForSEO all ship hosted servers. Check the plan gate first: verified August 2026, Ahrefs needs Lite at $129/month rather than the $29 Starter, and Nightwatch needs Professional at EUR 159/month and excludes the free trial.
15+ years of growing SaaS websites through SEO | Author, 200-Point Audit Checklist
Aditi has spent 15+ years helping SaaS companies scale organic traffic through technical SEO and content strategy. She is the author of the CrawlRaven 200-Point Audit checklist used by agencies and in-house teams to systematically improve search performance.