How to Add a Remote MCP Server to Claude
What a Remote MCP Server Actually Gives Claude
The Model Context Protocol (MCP) is an open standard that lets AI assistants like Claude call external tools, query live data, and take actions outside their training data. A 'remote' MCP server is simply one that runs somewhere other than your own machine — on a cloud host, a company's infrastructure, or a vendor's managed service — and exposes its tools over HTTP instead of requiring you to run a local process.
This distinction matters because Claude has historically supported two connection models: local servers launched from a config file using stdio, and remote servers reached over a URL, typically secured with OAuth. Remote servers are easier to adopt since there's nothing to install or keep running, and they're what most commercial MCP integrations — including document analysis tools — are built around.
Before You Start
You'll need a Claude plan that supports custom connectors (this is generally available on paid tiers, and increasingly on team/enterprise plans with admin controls), the URL of the MCP server you want to connect, and any credentials it requires — often just an OAuth login, sometimes an API key. If the server requires a subscription or account, set that up first so you have valid login details ready when Claude prompts for authentication.
Step-by-Step: Adding the Server in Claude
Open Claude in your browser or the desktop app and go to Settings. Look for a section labeled Connectors or Integrations — this is where Claude manages both first-party and custom tool connections.
Choose the option to add a custom connector (sometimes labeled 'Add custom integration' or 'Connect remote server'). You'll be prompted for the server's endpoint URL, which the provider of the MCP server should supply directly — it typically looks like a standard HTTPS URL ending in a path such as /mcp or /sse.
After entering the URL, Claude will attempt to establish a connection and discover the tools the server exposes. If the server requires authentication, you'll be redirected through an OAuth flow to log in and grant Claude permission to call its tools on your behalf. Once that completes, the connector shows up as active, and you can toggle which of its tools are available to Claude during conversations.
If you're using Claude Desktop and prefer the configuration-file approach for a locally-run server, note that this is a separate mechanism — remote servers don't need entries in claude_desktop_config.json. Mixing up the two is a common source of confusion, so confirm whether the server you're adding is meant to run locally or remotely before choosing a setup path.
Managing Permissions and Scope
Once connected, Claude will typically ask for confirmation before invoking a tool for the first time in a session, especially for actions that read or write data. You can usually review and revoke access to a connector at any time from the same Settings page, which is worth doing periodically if you're experimenting with several MCP servers and want to keep your active tool list clean.
For team or enterprise accounts, admins may control which connectors are approved for the organization, meaning individual users won't be able to add arbitrary remote servers without that server being allow-listed first. If you don't see the option to add a custom connector at all, this is the most likely reason.
Using the Tools Once Connected
After setup, you don't need any special syntax to use the new tools — just describe what you want in plain language, and Claude will decide when to invoke a connected tool based on the request and the tool's description. For example, if a connected server exposes a document summarization tool, asking Claude to 'summarize this contract and flag anything unusual' is enough for it to route the request appropriately, provided the relevant document is accessible to the tool (either uploaded, linked, or already available to the server).
It's good practice to test a new connector with a simple, low-stakes request first, so you can confirm the tool is returning what you expect before relying on it for anything important.
Troubleshooting Common Issues
If the connection fails immediately, double-check the URL for typos and confirm the server is actually online — remote MCP servers can go down like any other hosted service. If authentication loops without completing, clear cookies for the auth provider or try in a private browser window, since stale sessions are a frequent culprit.
If the connector connects but Claude doesn't seem to use the tool, check that the tool is toggled on in the connector's settings and that your request clearly matches what the tool is designed to do — vague prompts sometimes lead Claude to answer from general knowledge instead of invoking the tool.
A Practical Example: Document Analysis
One of the more immediately useful categories of remote MCP servers handles document work — summarizing long files, extracting key terms, or scanning contracts for risky clauses. If you're evaluating this kind of setup, Document Analysis MCP is a hosted MCP server offering summarize_document and find_red_flags tools that any MCP-compatible client, including Claude, can call directly after connecting via subscription — no local server to run or maintain.
Whatever remote MCP server you end up connecting, the setup pattern is the same: get the URL, add it as a custom connector, authenticate, and let Claude handle the rest based on how you phrase your requests.