Clients

Connect a Remote MCP Server to Cursor, Claude, and VS Code

How to connect remote MCP endpoints using official MCP remote-server guidance plus host-specific docs — Streamable HTTP, auth, and common failure modes.

10 min read Published July 14, 2026 Updated July 14, 2026

Official starting points

TopicOfficial resource
Remote connections (Claude-oriented guide)Connect to remote MCP servers
Local stdio connectionsConnect to local MCP servers
Transport rulesTransports — 2025-11-25
AuthAuthorization — 2025-11-25
Ecosystem client list (examples)What is MCP? — broad ecosystem support links Claude, ChatGPT, VS Code, Cursor, and others

Host UIs change frequently — always confirm syntax in that product’s current docs:

Before you configure

  1. Deployment running and reachable over HTTPS
  2. Transport is Streamable HTTP for new remotes (current standard) — or you knowingly support legacy HTTP+SSE per the backwards-compatibility guide
  3. Auth material ready (OAuth or HTTP bearer/API key as your server requires)
  4. Exact MCP endpoint path (spec example uses a path like /mcp)

MCPLambda product path: Connecting AI Clients.

Local vs remote config shapes

ModeTypical host configProtocol basis
Local stdiocommand + args (+ env)Client launches subprocess (stdio)
Remote HTTPServer URL (+ auth headers / OAuth)Streamable HTTP

Do not mix them for the same server entry.

Cursor

Use Cursor’s current MCP docs (cursor.com/docs/context/mcp). Remote servers are URL-based; local servers still use command/args. Reload MCP after config changes.

Claude Desktop / Claude remote

  • Local: claude_desktop_config.json with mcpServers command/args — see official build server and connect local guides (macOS path: ~/Library/Application Support/Claude/claude_desktop_config.json).
  • Remote: follow Connect to remote MCP servers and current Claude product connector UI — prefer product docs over outdated screenshots.

VS Code

Microsoft documents MCP server configuration for VS Code Copilot Chat: MCP servers in VS Code. Use that page for the live JSON schema and UI steps.

Auth

  • Prefer mechanisms aligned with the Authorization specification for OAuth-capable HTTP servers
  • Architecture overview notes HTTP bearer tokens, API keys, and custom headers as transport-layer options, with OAuth recommended for obtaining tokens: Architecture
  • Never commit tokens to git

Verify

  1. Client shows the server as connected
  2. tools/list (or host UI tool list) matches expectations
  3. One successful tool call
  4. Failures appear in server logs

Optional: MCP Inspector.

Common failures

IssueWhat to check
Wrong pathEndpoint must match server’s MCP path (often /mcp)
stdio config for remoteSwitch to URL-based remote config
401 / 403Auth scheme, token, Origin validation on server
Works in one client onlyProtocol version / transport support; try Streamable HTTP probe rules
Legacy remote onlyHTTP+SSE backwards-compatibility path in transports

Worked scenario: Three hosts, one URL

After deploy, Cursor works with a URL + header, Claude needs the connector OAuth flow, and VS Code uses its own MCP settings schema. Engineers assume the server is broken when only one host is misconfigured.

You publish a short matrix: server URL, transport (Streamable HTTP), auth method, and per-host deep links to current official docs. You forbid pasting long-lived tokens into git. Verification is always: connected → tools visible → one successful tool call → server logs show it.

Checklist for this topic

  • Confirm server Running and endpoint path
  • Use URL config for remote, command/args for stdio
  • Follow each host’s current official docs for JSON shape
  • Prefer HTTPS Streamable HTTP
  • Store tokens outside git
  • Verify with a real tool call, not only “connected”

Topic-specific failure modes

FailureLikely causeFix
Only one host worksHost-specific configPer-host docs matrix
401 everywhereAuth header/schemeAlign with server auth
Wrong pathMissing /mcpCopy exact deployment endpoint
Legacy SSE only clientTransport generationBack-compat or upgrade client

Transports · Auth patterns · Debugging

In practice: host matrix doc

Keep a living table in your wiki:

| Host | Config location | Remote fields | Auth notes | Last verified |

Update “last verified” when the vendor changes settings UI. Stale screenshots cause more outages than broken servers.

Smoke test script for humans

After any host config change:

  1. Reload/restart MCP in the host
  2. Confirm server status connected
  3. Invoke a harmless read tool
  4. Confirm server logs show the call
  5. Revoke a token in staging and confirm failure mode is clear

If step 4 never happens, you are debugging the wrong layer.

Takeaways

For Connect a Remote MCP Server to Cursor, Claude, and VS Code, remember three things:

  1. Be specific to this problem — the worked scenario “Three hosts, one URL” is the failure mode you should design against, not a generic outage narrative.
  2. Validate the same Deployment URL in each host you ship for — Cursor, Claude, and VS Code each have different config surfaces.
  3. Treat host config drift as a release risk: document the exact JSON/UI steps and re-test after host updates.

If you only remember one habit: one green host is not three green hosts — smoke each client after every server deploy.

Sources

Next steps

Debugging MCP in production · Auth patterns

FAQs

Frequently Asked Questions

  • What URL do I paste into the client?

    The server’s MCP HTTP endpoint. The Streamable HTTP specification requires a single MCP endpoint path that supports POST and GET (example shape https://example.com/mcp). On MCPLambda this is often https://<deployment>.mcplambda.io/mcp — confirm in your deployment UI.

  • Do I still need command/args for remote servers?

    No. Official local setups use command/args for stdio subprocesses. Remote Streamable HTTP servers are reached by URL (plus auth as required). See Connect to remote MCP servers on modelcontextprotocol.io.