# How to Add the RankX AI MCP Server to Claude > Source: https://rankxai.com/blog/connect-rankx-to-claude ยท Last updated: 2026-08-21 Claude connects to RankX AI over MCP in under a minute. In Claude Desktop or Claude Web, add a custom connector pointing at the RankX AI endpoint and sign in with OAuth. In Claude Code, run one claude mcp add command with either OAuth or a personal access token. An account owner approves the connection once. ## What does adding an MCP server to Claude do? Adding an MCP server gives Claude tools: functions the assistant can call to read your data sources and perform actions on external systems, live, mid-conversation. Add the RankX AI MCP server and your visibility data, rankings, keyword research and site audits become tools that Claude can use the moment you ask, which is the whole reason to bother: an assistant that can interact with external systems answers with your numbers instead of general advice. You can add MCP servers to Claude three ways, and this step-by-step guide covers each for the RankX AI server specifically: the Connectors panel in Claude Desktop and Claude Web, one command in the Claude Code CLI, and a token in an editor's MCP configuration. The integration takes under a minute on any of them. ## How do you connect Claude Desktop or Claude Web? Claude Desktop and Claude Web connect to RankX AI through the Connectors panel, with OAuth. In Claude, open Settings, then Connectors, then Add custom connector, and paste the endpoint URL: `https://app.rankxai.com/api/mcp`. Leave the client id and client secret empty; RankX AI registers your client automatically, which is what the protocol expects for a public client, and filling those fields in with something invented will fail. There is no configuration file to edit and nothing to restart; if a guide tells you to restart Claude Desktop after editing a JSON file, it is describing local MCP servers, the other kind, covered below. Claude then sends you to RankX AI to sign in. The consent screen names every capability the connection is asking for, and only an account owner can approve it, a deliberate boundary because this credential can spend credits and edit a live site. Approve once and it stays connected. One warning worth its own sentence: do not try to do this by editing Claude Desktop's config file. That file starts local programs only, an HTTP entry in it is silently ignored, and the failure looks like the server being down. The [Claude Desktop page](/docs/mcp/connect/claude-desktop) documents the trap and the bridge around it. ## How do you add the MCP server to Claude Code? The fastest way to set up MCP with Claude Code is one command in the Claude Code CLI. For OAuth, run `claude mcp add --transport http rankxai https://app.rankxai.com/api/mcp` and a browser opens for you to sign in and approve. For a token, append a header: the same command plus `--header "Authorization: Bearer rxai_your_token_here"`, with the token created first in RankX AI's MCP settings. Then start Claude Code and verify with `claude mcp list`, where the new server should show as connected, and ask for something real: using RankX AI, list my websites and show me the AI visibility for the first one over the last 30 days. Claude asks permission before its first tool calls, which is the moment you know the MCP server reached Claude Code properly. The assistant should call `list_projects` first. Every other tool takes a website id that call returns, so it is the entry point rather than a formality. By default `claude mcp add` writes to your user configuration, making the connection available in every project; to configure it for one repository instead, add it to a `.mcp.json` in that project directory with an environment variable reference for the credential, which is how MCP installation scopes work. The [Claude Code page](/docs/mcp/connect/claude-code) has both forms ready to paste. ## What about local MCP servers? RankX AI needs no local install, but the other kind is worth recognising. Local MCP servers run on your own machine as stdio servers, subprocesses Claude starts itself, and exist for jobs a hosted service cannot reach: filesystem access, access to specific directories, driving software you run locally. Those are the setups that involve editing a configuration file and restarting; a working setup often runs both kinds side by side, GitHub's remote server and a local filesystem server next to RankX AI, and Claude treats every connected server the same way in conversation. ## Should you use OAuth or a token? OAuth, when a person is at the keyboard: no secret in your shell history, and revoking it is a click. A personal access token, when the connection has to survive without one, or when you want less than everything. OAuth grants the full capability set in one decision; a token picks its scopes at creation, which raw API keys never did, so it is the way to mint a read-only credential that can report but never write or spend, or a client-scoped credential that reaches one client's websites only. Scripts and automation cannot use OAuth at all: RankX AI's authorisation server implements the authorisation-code flow with PKCE and nothing else, so every browser-issued credential requires a human present, and a cron job has no browser. Both credential types resolve to the same authorisation model on the server, so nothing else about the connection differs. The [authentication documentation](/docs/mcp/authentication) covers scopes, client-scoped tokens and revocation in full. ## What can Claude do once connected? The connection exposes 66 tools, from prompt visibility and AI Overview citations through keyword research, site audits and Search Console reads, each declaring the scope it requires. They sit alongside whatever other MCP integrations you already run, GitHub for repositories, a local filesystem server, and Claude composes them freely. On top of the RankX AI tools sit six Agent Skills, written workflows that run whole jobs: the one to start with is the visibility audit, and [running it end to end](/blog/running-your-seo-from-claude-and-chatgpt) is this cluster's walkthrough. [What Claude Skills are](/blog/what-are-claude-skills), and the [tool reference](/docs/mcp/tool-reference) with all 66 grouped, cover the rest of the surface. ## Troubleshooting: what usually goes wrong? - The Claude Desktop config file. It cannot express an HTTP connection, and the failure is silent. Use the Connectors panel. - A missing tool. If a tool you expected is not in the list, the credential does not carry that tool's scope; a tool outside the connection's scopes is omitted entirely rather than shown and denied. Issue a token with the right scopes, or reconnect with OAuth. - A non-owner trying to approve. The consent screen tells you rather than half-connecting; get the account owner to approve once. - A token in a committed file. Revoke it in RankX AI's MCP settings and issue a new one. Beyond those four, the [troubleshooting page](/docs/mcp/troubleshooting) works through the rarer failures with the debug steps for each. Once the connection is live, RankX AI stays a remote MCP server: nothing runs on your machine, credentials stay server-side, and [why remote beats local](/blog/remote-mcp-servers) is the architecture behind that choice. ## Sources - [Anthropic, Introducing the Model Context Protocol, November 2024](https://www.anthropic.com/news/model-context-protocol), checked 20 Aug 2026 - [Model Context Protocol specification, authorization (OAuth 2.1 with PKCE)](https://modelcontextprotocol.io/specification), checked 20 Aug 2026 ## Does connecting cost anything? The connection is free and reading is always free. The handful of tools that spend credits state their current price in their own description when the client lists them, so the assistant knows what a call costs before making one, and the server instructs it to check your balance and confirm with you before proposing work that spends. ## Can I connect ChatGPT and the code editors too? Yes. ChatGPT connects through its own connector settings with the same browser sign-in, and Cursor, VS Code and Windsurf take a personal access token in each editor's config file. Scripts and CI use a token as the only supported path, because OAuth here always requires a human at a browser. The documentation has a page per client with the exact block to paste. ## Is a token safe to put in a repository? No. A committed token is a leaked credential, even in a private repository. Commit the configuration file with an environment variable reference instead, and keep the token itself in your environment. If a token does leak, revoke it in RankX AI's MCP settings and issue a new one; the swap is zero downtime.