mcp-remote

Connect an MCP Client that only supports local (stdio) servers to a Remote MCP Server, with auth support:

Why is this necessary?

So far, the majority of MCP servers in the wild are installed locally, using the stdio transport. This has some benefits: both the client and the server can implicitly trust each other as the user has granted them both permission to run. Adding secrets like API keys can be done using environment variables and never leave your machine. And building on npx and uvx has allowed users to avoid explicit install steps, too.

But there's a reason most software that could be moved to the web did get moved to the web: it's so much easier to find and fix bugs & iterate on new features when you can push updates to all your users with a single deploy.

With the latest MCP Authorization specification, we now have a secure way of sharing our MCP servers with the world without running code on user's laptops. Or at least, you would, if all the popular MCP clients supported it yet. Most are stdio-only, and those that do support HTTP+SSE don't yet support the OAuth flows required.

That's where mcp-remote comes in. As soon as your chosen MCP client supports remote, authorized servers, you can remove it. Until that time, drop in this one liner and dress for the MCP clients you want!

Usage

All the most popular MCP clients (Claude Desktop, Cursor & Windsurf) use the following config format:

{
  "mcpServers": {
    "remote-example": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse"
      ]
    }
  }
}

Custom Headers

To bypass authentication, or to emit custom headers on all requests to your remote server, pass --header CLI arguments:

{
  "mcpServers": {
    "remote-example": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--header",
        "Authorization: Bearer ${AUTH_TOKEN}"
      ],
      "env": {
        "AUTH_TOKEN": "..."
      }
    },
  }
}

Note: Cursor, Codex-Cli and Claude Desktop (Windows) have a bug where spaces inside args aren't escaped when it invokes npx, which ends up mangling these values. You can work around it using:

{
  // rest of config...
  "args": [
    "mcp-remote",
    "https://remote.mcp.server/sse",
    "--header",
    "Authorization:${AUTH_HEADER}" // note no spaces around ':'
  ],
  "env": {
    "AUTH_HEADER": "Bearer <auth-token>" // spaces OK in env vars
  }
},

To keep a credential out of the process arguments — where any other user on the machine can read it from the process list — put the headers in a file instead and pass --header-file. One Name: value per line; # starts a comment.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--header-file",
        "/path/to/headers.txt"
      ]

# credentials for the example server
Authorization: Bearer my-token
X-Custom-Header: custom-value

A file that cannot be read is an error rather than a warning, so a mistyped path fails immediately instead of sending the request unauthenticated.

Multiple Instances

To run multiple instances of the same remote server with different configurations (e.g., different Atlassian tenants), use the --resource flag to isolate OAuth sessions:

{
  "mcpServers": {
    "atlassian_tenant1": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://mcp.atlassian.com/v1/sse",
        "--resource",
        "https://tenant1.atlassian.net/"
      ]
    },
    "atlassian_tenant2": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://mcp.atlassian.com/v1/sse",
        "--resource",
        "https://tenant2.atlassian.net/"
      ]
    }
  }
}

Each unique combination of server URL, resource, custom headers, and --authorize-param values will maintain separate OAuth sessions and token storage.

The --resource value is sent as the RFC 8707 resource indicator on the authorization, token and refresh requests alike, so they always agree.

Extra authorization parameters

Some authorization servers require parameters of their own on the authorize call. Pass each as --authorize-param key=value, repeating the flag as needed:

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/mcp",
        "--authorize-param",
        "access_type=offline",
        "--authorize-param",
        "prompt=consent"
      ]

Those two are what Google wants before it will part with a refresh token — it does not recognise the offline_access scope. Auth0 wants audience=https://your-api to issue a JWT rather than an opaque token. [email protected] is also common.

These apply to the authorization request only. resource is the exception: RFC 8707 wants the same value on the token and refresh requests too, and only --resource puts it there. Parameters the flow derives per request — state, code_challenge, client_id, redirect_uri, response_type — are refused, because a value that disagrees with the real one surfaces as an opaque server error.

Changing these starts a new sign-in, since a parameter like audience decides which API the token is for and a token issued for one is not valid for another.

Some authorization servers reject the resource parameter outright — Microsoft Entra ID v2 answers AADSTS9010010, for example. Pass --disable-resource-parameter to omit it entirely:

{
  "mcpServers": {
    "entra-example": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://remote.mcp.server/mcp",
        "--disable-resource-parameter"
      ]
    }
  }
}

Flags

• If npx is producing errors, consider adding -y as the first argument to auto-accept the installation of the mcp-remote package.

      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://remote.mcp.server/sse"
      ]

• To force npx to always check for an updated version of mcp-remote, add the @latest flag:

      "args": [
        "mcp-remote@latest",
        "https://remote.mcp.server/sse"
      ]

• To change which port mcp-remote listens for an OAuth redirect, add an additional argument after the server URL. By default the port is derived from the server URL, so every server gets a stable port of its own somewhere in 3335-49150, and mcp-remote walks up to 8 ports from there if it finds one taken. A port you pass explicitly is used as-is: it implies a redirect_uri the authorization server has already been given, so mcp-remote fails rather than quietly moving to a different one. --static-oauth-client-info pins the port the same way, for the same reason.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "9696"
      ]

• To change which host mcp-remote registers as the OAuth callback URL (by default localhost), add the --host flag.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--host",
        "127.0.0.1"
      ]

• To change the path mcp-remote serves the OAuth callback on (by default /oauth/callback), add the --callback-path flag. The path must start with /, and /wait-for-auth is reserved.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--callback-path",
        "/custom/callback"
      ]

• To allow HTTP connections in trusted private networks, add the --allow-http flag. Note: This should only be used in secure private networks where traffic cannot be intercepted.

      "args": [
        "mcp-remote",
        "http://internal-service.vpc/sse",
        "--allow-http"
      ]

• To enable detailed debugging logs, add the --debug flag. This will write verbose logs to ~/.mcp-auth/{server_hash}_debug.log with timestamps and detailed information about the auth process, connections, and token refreshing.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--debug"
      ]

• To suppress default logs, add the --silent flag. This will prevent logs from being emitted, except in the case where --debug is also passed.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--silent"
      ]

• To enable an outbound HTTP(S) proxy for mcp-remote, add the --enable-proxy flag. When enabled, mcp-remote will use the proxy settings from common environment variables (for example HTTP_PROXY, HTTPS_PROXY, and NO_PROXY).

    "args": [
      "mcp-remote",
      "https://remote.mcp.server/sse",
      "--enable-proxy"
    ],
    "env": {
      "HTTPS_PROXY": "http://127.0.0.1:3128",
      "NO_PROXY": "localhost,127.0.0.1"
    }

• To ignore specific tools from the remote server, add the --ignore-tool flag. This will filter out tools matching the specified patterns from both tools/list responses and block tools/call requests. Supports wildcard patterns with *.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--ignore-tool",
        "delete*",
        "--ignore-tool",
        "remove*"
      ]

You can specify multiple --ignore-tool flags to ignore different patterns. Examples:

• delete* - ignores all tools starting with "delete" (e.g., deleteTask, deleteUser)
• *account - ignores all tools ending with "account" (e.g., getAccount, updateAccount)
• exactTool - ignores only the tool named exactly "exactTool"

• To change the timeout for the OAuth callback (by default 30 seconds), add the --auth-timeout flag with a value in seconds. This is useful if the authentication process on the server side takes a long time.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--auth-timeout",
        "60"
      ]

• To change the network timeouts, add --connect-timeout, --headers-timeout or --body-timeout, each with a value in seconds. These apply to every outbound request, including the OAuth ones.

  • --connect-timeout bounds establishing the TCP connection (default 10). Lower it to fail faster on an unreachable server.
  • --headers-timeout bounds waiting for response headers (default 300).
  • --body-timeout bounds a gap between chunks of a response body (default 300). This is the one that closes an idle SSE stream after five minutes; pass 0 to disable it for servers that push infrequently.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--connect-timeout",
        "30",
        "--body-timeout",
        "0"
      ]

• To stop an idle connection being dropped, add the --keep-alive flag. The proxy then sends a ping every 30 seconds, which is enough traffic to keep a server — or a load balancer in front of one — from reaping a session that has been quiet for a few minutes. Use --ping-interval with a value in seconds to change the period; setting it turns keep-alive on, so the two flags are only both needed when you want the default period spelled out.

  This is the opposite end of the problem from --body-timeout: that one governs how long we wait, whereas this keeps the other side from hanging up.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--keep-alive",
        "--ping-interval",
        "60"
      ]

• To connect over IPv4 only, add the --ipv4 flag. Useful when a hostname resolves to both IPv4 and IPv6 addresses but the IPv6 routes silently black-hole rather than being refused — connection attempts then time out instead of failing over, and the request never completes.

      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse",
        "--ipv4"
      ]

Transport Strategies

MCP Remote supports different transport strategies when connecting to an MCP server. This allows you to control whether it uses Server-Sent Events (SSE) or HTTP transport, and in what order it tries them.

Specify the transport strategy with the --transport flag:

npx mcp-remote https://example.remote/server --transport sse-only

Available Strategies:

• http-first (default): Tries HTTP transport first, falls back to SSE if HTTP fails with a 404 error
• sse-first: Tries SSE transport first, falls back to HTTP if SSE fails with a 405 error
• http-only: Only uses HTTP transport, fails if the server doesn't support it
• sse-only: Only uses SSE transport, fails if the server doesn't support it

Protocol Eras (2026-07-28 servers)

The 2026-07-28 revision of MCP retired the initialize handshake and the session behind it. A server that implements it serves stateless requests that each carry their own protocol version and capabilities, and advertises itself through server/discover instead of answering a handshake.

Most desktop hosts still speak the 2025-11-25 era. The specification's compatibility matrix puts that pair — legacy client, modern server — in the one cell that simply fails, because a legacy client has no way to fall forward. The fix it names is a dual-era client, which is what mcp-remote becomes with --protocol auto:

npx mcp-remote https://example.remote/mcp --protocol auto

Available modes:

• legacy (default): sends the client's initialize straight through, exactly as every earlier release did. Nothing is probed.
• auto: spends one server/discover on the first handshake. If a 2026-07-28 server answers, mcp-remote answers the local client's handshake itself and rewrites every later request into the modern era — per-request _meta, matching MCP-Protocol-Version header. If anything else answers, the handshake goes out untouched and nothing changes.

It is off by default because every server in the wild today is a legacy one, and the probe is a round trip those connections do not need.

The two modern surfaces with no 2025-11-25 equivalent are bridged as well:

• Change notifications. The modern era only sends notifications/tools/list_changed and friends down a subscriptions/listen stream the client opens. A 2025-era client never opens one, so mcp-remote opens it on the client's behalf — asking for exactly the notifications the server said it can send — and passes each one on in the shape that era expects.
• Multi-round-trip requests. When a server answers with input_required, asking for sampling, elicitation or roots mid-request, mcp-remote unpacks the embedded questions and puts them to the client as the ordinary server-initiated requests it already understands, then retries the original request with the answers and a byte-exact echo of the server's requestState. The client is never told any of this happened: it is still waiting on the one request it sent, and that is what it is answered with. Capped at 10 rounds.

A question this proxy cannot put to a 2025-era client — anything outside sampling, elicitation and roots, or a URL-mode elicitation, or one the client never declared it supports — is reported as an error rather than dropped.

Three methods the modern era retired are honoured by mcp-remote itself rather than forwarded to a server that no longer has them, because the capabilities the client is handed still advertise them:

• resources/subscribe / resources/unsubscribe become entries on the subscriptions/listen stream, which is reopened whenever the set changes.
• logging/setLevel is recorded and carried as the per-request io.modelcontextprotocol/logLevel that replaced it — without which a modern server sends no logs at all.

Static OAuth Client Metadata

MCP Remote supports providing static OAuth client metadata instead of using the mcp-remote defaults. This is useful when connecting to OAuth servers that expect specific client/software IDs or scopes.

Provide the client metadata as a JSON string or as a @ prefixed filepath with the --static-oauth-client-metadata flag:

npx mcp-remote https://example.remote/server --static-oauth-client-metadata '{ "scope": "space separated scopes" }'
# uses node readfile, so you probably want to use absolute paths if you're not sure what the cwd is
npx mcp-remote https://example.remote/server --static-oauth-client-metadata '@/Users/username/Library/Application Support/Claude/oauth_client_metadata.json'

What you provide is merged over the defaults, so this is also how you pin a value mcp-remote would otherwise negotiate. token_endpoint_auth_method is picked from the authorization server's token_endpoint_auth_methods_supported: none when the server accepts public clients, otherwise client_secret_post, otherwise client_secret_basic. Override it when the server needs something else:

npx mcp-remote https://example.remote/server --static-oauth-client-metadata '{ "token_endpoint_auth_method": "client_secret_post" }'

Static OAuth Client Information

Per the spec, servers are encouraged but not required to support OAuth dynamic client registration.

For these servers, MCP Remote supports providing static OAuth client information instead. This is useful when connecting to OAuth servers that require pre-registered clients.

Provide the client metadata as a JSON string or as a @ prefixed filepath with the --static-oauth-client-info flag:

export MCP_REMOTE_CLIENT_ID=xxx
export MCP_REMOTE_CLIENT_SECRET=yyy
npx mcp-remote https://example.remote/server --static-oauth-client-info "{ \"client_id\": \"$MCP_REMOTE_CLIENT_ID\", \"client_secret\": \"$MCP_REMOTE_CLIENT_SECRET\" }"
# uses node readfile, so you probably want to use absolute paths if you're not sure what the cwd is
npx mcp-remote https://example.remote/server --static-oauth-client-info '@/Users/username/Library/Application Support/Claude/oauth_client_info.json'

Client ID Metadata Documents

SEP-991 lets an authorization server accept an HTTPS URL as the client_id, where that URL serves a JSON document describing the client. Servers that support it advertise "client_id_metadata_document_supported": true in their authorization server metadata, and no dynamic registration is needed.

Point mcp-remote at your document with --client-metadata-url:

npx mcp-remote https://example.remote/server --client-metadata-url https://client.example.com/.well-known/oauth-client-metadata

The URL must use HTTPS and have a path — a bare origin is rejected. If the server does not advertise support, mcp-remote registers dynamically as usual, so the flag is safe to leave in place.

The document you host has to list the redirect_uri mcp-remote will send, which contains the OAuth callback port. Passing this flag makes that port strict: mcp-remote uses the port derived from the server URL (or the one you pass explicitly) and fails rather than quietly moving to another one, so the redirect_uris in your document stay correct. Run once to see the port it picked, or choose it yourself:

npx mcp-remote https://example.remote/server 3334 --client-metadata-url https://client.example.com/.well-known/oauth-client-metadata

Signing In Without a User

Some servers are not protected on behalf of a person at all — an internal MCP server reached by a scheduled job, say, where there is no user to consent and no browser to consent in. For those, --client-credentials uses the OAuth Client Credentials grant: the client presents its own credentials and is issued a token for itself.

npx mcp-remote https://example.remote/mcp \
  --client-credentials \
  --static-oauth-client-info '{"client_id":"my-client","client_secret":"${MCP_CLIENT_SECRET}"}'

The credentials come from --static-oauth-client-info, which accepts inline JSON or @path/to/file.json. ${ENV_VAR} placeholders are expanded from the environment in both forms, so the secret need not sit in a command line where every other process on the machine can read it. Nothing logs the expanded value.

The token endpoint, the scope and the RFC 8707 resource indicator are discovered and applied the same way as for every other flow, and client_secret_basic or client_secret_post is chosen from what the authorization server advertises. No refresh token is involved: the credentials are the durable thing, so an expired token is replaced by asking for another the same way the first was obtained.

If an internal server does not publish RFC 9728 or RFC 8414/OIDC discovery metadata, provide its known token endpoint explicitly. This skips discovery and is accepted only with --client-credentials:

npx mcp-remote https://example.remote/mcp \
  --client-credentials \
  --token-endpoint https://auth.example.com/oauth/token \
  --static-oauth-client-info '{"client_id":"my-client","client_secret":"${MCP_CLIENT_SECRET}"}' \
  --static-oauth-client-metadata '{"scope":"mcp.read mcp.write","token_endpoint_auth_method":"client_secret_basic"}'

The explicit endpoint must use HTTPS, except for an HTTP loopback endpoint. Credentials and URL fragments are refused in the endpoint URL, and changing the endpoint selects a separate token cache. Discovery is also skipped on cold startup and 401 retries. Tokens are renewed before expiry using the same client credentials; if renewal fails, the current token is sent until the server refuses it. With an explicit endpoint, scope is omitted unless configured with --static-oauth-client-metadata or named in the server's 401 challenge, and resource is omitted unless set with --resource.

Signing In Without a Browser

The default flow needs a browser on this machine and a loopback port to redirect back to — which a cron job, an SSH session or a container does not have. If your authorization server supports the OAuth Device Authorization Grant, --device-code moves the browser to whatever machine you are actually sitting at:

npx mcp-remote https://example.remote/server --device-code

mcp-remote prints a short code and a URL, then polls until you approve it:

To authorize this client, visit:
  https://auth.example.com/activate

And enter the code: WDJB-MJHT

Waiting for approval...

The output goes to stderr, which MCP clients capture into their own logs — so in a headless run, that log is where you read the code. This only has to happen once: the refresh token that comes back is stored like any other, and later runs are non-interactive.

No callback server is started and no port is bound, so this is also the flow to use when the loopback port is unavailable.

The server has to advertise device_authorization_endpoint in its authorization server metadata; mcp-remote fails with a clear message rather than falling back to a browser that isn't there. Note that servers offering this grant often expect a pre-registered client — pair it with --static-oauth-client-info if dynamic registration is refused.

Load Balancer Session Stickiness

MCP sessions live on one backend, so a server behind a load balancer needs every request from a client to reach the same node. AWS ALB, Azure Load Balancer and others do this with a cookie the client is expected to send back.

mcp-remote keeps cookies the MCP server sets and replays them on later requests, including from the SSE stream onto the POSTs that follow it. Nothing to configure. Cookies are held in memory for the life of the process, never written to disk, and only ever sent back to the exact origin that set them — Domain is ignored, so nothing travels to another host.

A Cookie you pass yourself with --header always wins.

To turn it off:

npx mcp-remote https://example.remote/server --disable-cookies

Using the ID Token as the Bearer Credential

By default mcp-remote sends the OAuth access token, which says what the caller is allowed to do. Some servers instead verify who the caller is: they validate an OIDC ID token against a JWKS endpoint and read identity claims such as sub and email from it. AWS Cognito in front of Bedrock AgentCore works this way, and rejects the access token outright.

Pass --use-id-token to send the ID token instead:

npx mcp-remote https://example.remote/server --use-id-token

Only the credential presented to the MCP server changes — the refresh token and the renewal flow are untouched. Renewal follows the ID token's own exp claim rather than the access token's lifetime, since the ID token is what actually goes on the wire.

An ID token is only issued when openid is among the requested scopes. If your server does not advertise it, ask for it explicitly:

npx mcp-remote https://example.remote/server --use-id-token --static-oauth-client-metadata '{ "scope": "openid email" }'

Claude Desktop

Official Docs

In order to add an MCP server to Claude Desktop you need to edit the configuration file located at:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

If it does not exist yet, you may need to enable it under Settings > Developer.

Restart Claude Desktop to pick up the changes in the configuration file. Upon restarting, you should see a hammer icon in the bottom right corner of the input box.

Cursor

Official Docs. The configuration file is located at ~/.cursor/mcp.json.

As of version 0.48.0, Cursor supports unauthed SSE servers directly. If your MCP server is using the official MCP OAuth authorization protocol, you still need to add a "command" server and call mcp-remote.

Windsurf

Official Docs. The configuration file is located at ~/.codeium/windsurf/mcp_config.json.

Building Remote MCP Servers

For instructions on building & deploying remote MCP servers, including acting as a valid OAuth client, see the following resources:

• https://developers.cloudflare.com/agents/guides/remote-mcp-server/

In particular, see:

• https://github.com/cloudflare/workers-oauth-provider for defining an MCP-comlpiant OAuth server in Cloudflare Workers
• https://github.com/cloudflare/agents/tree/main/examples/mcp for defining an McpAgent using the agents framework.

For more information about testing these servers, see also:

• https://developers.cloudflare.com/agents/guides/test-remote-mcp-server/

Know of more resources you'd like to share? Please add them to this Readme and send a PR!

Troubleshooting

Clear your ~/.mcp-auth directory

mcp-remote stores all the credential information inside ~/.mcp-auth (or wherever your MCP_REMOTE_CONFIG_DIR points to). If you're having persistent issues, try running:

rm -rf ~/.mcp-auth

Then restarting your MCP client.

Credentials are stored under mcp-remote-v1, which names the layout of the store rather than the version of the package, so upgrading mcp-remote no longer signs you out. Releases before this change kept a separate directory per version — if you have ~/.mcp-auth/mcp-remote-0.x.y directories left over, they hold old tokens and can be deleted.

Check your Node version

Make sure that the version of Node you have installed is 18 or higher. Claude Desktop will use your system version of Node, even if you have a newer version installed elsewhere.

Restart Claude

When modifying claude_desktop_config.json it can helpful to completely restart Claude

VPN Certs

You may run into issues if you are behind a VPN, you can try setting the NODE_EXTRA_CA_CERTS environment variable to point to the CA certificate file. If using claude_desktop_config.json, this might look like:

{
 "mcpServers": {
    "remote-example": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://remote.mcp.server/sse"
      ],
      "env": {
        "NODE_EXTRA_CA_CERTS": "{your CA certificate file path}.pem"
      }
    }
  }
}

Check the logs

• Follow Claude Desktop logs in real-time
• MacOS / Linux:
  tail -n 20 -F ~/Library/Logs/Claude/mcp*.log
• For bash on WSL:
  tail -n 20 -f "C:\Users\YourUsername\AppData\Local\Claude\Logs\mcp.log"
• Powershell:
  Get-Content "C:\Users\YourUsername\AppData\Local\Claude\Logs\mcp.log" -Wait -Tail 20

Debugging

Debug Logs

For troubleshooting complex issues, especially with token refreshing or authentication problems, use the --debug flag:

"args": [
  "mcp-remote",
  "https://remote.mcp.server/sse",
  "--debug"
]

This creates detailed logs in ~/.mcp-auth/{server_hash}_debug.log with timestamps and complete information about every step of the connection and authentication process. When you find issues with token refreshing, laptop sleep/resume issues, or auth problems, provide these logs when seeking support.

Authentication Errors

If you encounter the following error, returned by the /callback URL:

Authentication Error
Token exchange failed: HTTP 400

You can run rm -rf ~/.mcp-auth to clear any locally stored state and tokens.

"Client" mode

Run the following on the command line (not from an MCP server):

npx -p mcp-remote@latest mcp-remote-client https://remote.mcp.server/sse

This will run through the entire authorization flow and attempt to list the tools & resources at the remote URL. Try this after running rm -rf ~/.mcp-auth to see if stale credentials are your problem, otherwise hopefully the issue will be more obvious in these logs than those in your MCP client.

Acknowledgments

Glen Maddern is the original author of mcp-remote. He built mcp-remote into one of the most popular building blocks in the MCP ecosystem.