rust-sdk/docs/OAUTH_SUPPORT.md

Model Context Protocol OAuth Authorization

This document describes the OAuth 2.1 authorization implementation for Model Context Protocol (MCP), following the MCP Authorization Specification.

Features

• Full support for OAuth 2.1 authorization flow with PKCE (S256)
• RFC 8707 resource parameter binding
• Protected Resource Metadata discovery (RFC 9728)
• Authorization Server Metadata discovery (RFC 8414 + OpenID Connect)
• Dynamic client registration (RFC 7591)
• Client ID Metadata Documents (CIMD) (SEP-991 / Client ID Metadata Documents )
• Scope selection from WWW-Authenticate, Protected Resource Metadata, and AS metadata
• Scope upgrade on 403 insufficient_scope (SEP-835)
• Automatic token refresh
• Authorized HTTP Client implementation
• Injectable OAuth HTTP client for custom network environments
• Opt-in EMA/XAA refresh-token and ID-JAG exchanges for registered public and confidential clients

Usage Guide

1. Enable Features

Enable the auth feature in Cargo.toml:

[dependencies]
rmcp = { version = "0.1", features = ["auth", "transport-streamable-http-client-reqwest"] }

2. Configure OAuth network requests

OAuth makes several HTTP requests before the MCP transport is connected: protected-resource discovery, authorization-server discovery, dynamic client registration, authorization-code exchange, token refresh, and client credentials exchange. When no OAuth HTTP client is provided, the SDK sends those requests with an internally-created reqwest::Client.

If you only need to customize reqwest behavior, pass a configured reqwest::Client to OAuthState::new. This preserves the caller-provided reqwest configuration across OAuth operations, including token requests.

let default_headers = reqwest::header::HeaderMap::new();
let oauth_http_client = reqwest::Client::builder()
    .timeout(std::time::Duration::from_secs(60))
    .default_headers(default_headers)
    .build()?;

let mut oauth_state = OAuthState::new(&server_url, Some(oauth_http_client))
    .await
    .context("Failed to initialize oauth state machine")?;

This is useful for proxy, TLS root, connector, timeout, and default-header configuration while staying within reqwest. The redirect behavior is the behavior of the provided reqwest client, so configure that client accordingly. This OAuth HTTP client is separate from the reqwest::Client later passed to AuthClient::new, which is used for the authorized MCP transport after tokens have been obtained.

If OAuth requests must run outside reqwest, implement OAuthHttpClient and use OAuthState::new_with_oauth_http_client. The SDK passes each OAuth request to your implementation with the raw HTTP request, a suggested timeout, and an OAuthHttpRedirectPolicy. OAuthHttpClientFuture returns OAuthHttpClientError, so implementations can propagate their native error types with ? without flattening their source chains into strings.

use std::sync::Arc;

use rmcp::transport::{
    OAuthHttpClient, OAuthHttpClientFuture, OAuthHttpRedirectPolicy,
    OAuthHttpRequest, OAuthState,
};

struct MyOAuthHttpClient;

impl OAuthHttpClient for MyOAuthHttpClient {
    fn execute(&self, request: OAuthHttpRequest) -> OAuthHttpClientFuture<'_> {
        Box::pin(async move {
            match request.redirect_policy {
                OAuthHttpRedirectPolicy::Follow => {
                    // Follow redirects according to your HTTP environment.
                }
                OAuthHttpRedirectPolicy::Stop => {
                    // Return redirect responses without following them.
                }
                _ => {
                    // Future redirect policies may be added.
                }
            }

            // Convert `request.request` into your HTTP stack's request type,
            // execute it, then convert the response back into the expected
            // OAuth HTTP response type.
            let response = todo!("send OAuth request");
            Ok(response)
        })
    }
}

let mut oauth_state = OAuthState::new_with_oauth_http_client(
    &server_url,
    Arc::new(MyOAuthHttpClient),
)
.await?;

Use this path when OAuth traffic must go through a browser fetch API, a remote execution environment, a company gateway, a test fake, or any other non-reqwest transport.

Inspect discovery provenance directly

Most applications can use OAuthState without calling metadata discovery directly. When using AuthorizationManager, resolve_metadata() returns both the metadata and how it was obtained. A client that supports the 2025-03-26 default-endpoint fallback can continue with synthesized metadata, while a client that requires server-published metadata should reject that result:

use rmcp::transport::auth::{AuthorizationManager, AuthorizationMetadataSource};

async fn configure_metadata(
    manager: &mut AuthorizationManager,
    allow_legacy_endpoint_fallback: bool,
) -> anyhow::Result<()> {
    let resolution = manager.resolve_metadata().await?;

    if resolution.source == AuthorizationMetadataSource::LegacyEndpointFallback {
        if !allow_legacy_endpoint_fallback {
            anyhow::bail!("the server did not publish OAuth metadata");
        }

        tracing::warn!(
            "the server did not publish OAuth metadata; using the 2025-03-26 fallback endpoints"
        );
    }

    manager.set_metadata(resolution.metadata);
    Ok(())
}

ProtectedResourceMetadata and AuthorizationServerMetadata indicate server-published metadata, so clients can proceed with the returned metadata. LegacyEndpointFallback indicates endpoints synthesized for compatibility with the 2025-03-26 MCP specification. Clients should proceed only when they intentionally support that legacy behavior; clients using discovery as an OAuth capability check should treat it as unsupported.

Applications using OAuthState do not need to handle these sources directly: the state machine resolves metadata internally and retains the legacy fallback. Low-level AuthorizationManager users can use AuthorizationMetadataSource::is_discovered() when they only need to distinguish server-published metadata from synthesized metadata.

3. Start authorization with OAuthState

The OAuthState state machine manages the full authorization lifecycle. start_authorization accepts an AuthorizationRequest describing the client identity material you have available, and selects a client registration mechanism following the spec's priority order:

1. Pre-registered client information (with_preregistered_client), when the client already holds a client_id issued out of band
2. Client ID Metadata Documents (SEP-991, with_client_metadata_url), when the authorization server advertises client_id_metadata_document_supported
3. Dynamic Client Registration, as a fallback when the authorization server advertises a registration_endpoint

When no scopes are provided, the SDK automatically selects scopes from the server's WWW-Authenticate header, Protected Resource Metadata, or AS metadata.

use rmcp::transport::auth::AuthorizationRequest;

// start authorization - pass no scopes to let the SDK auto-select
oauth_state
    .start_authorization(
        AuthorizationRequest::new(MCP_REDIRECT_URI).with_client_name("My MCP Client"),
    )
    .await
    .context("Failed to start authorization")?;

If you know the scopes you need, you can still pass them explicitly:

oauth_state
    .start_authorization(
        AuthorizationRequest::new(MCP_REDIRECT_URI)
            .with_scopes(["mcp", "profile"])
            .with_client_name("My MCP Client"),
    )
    .await
    .context("Failed to start authorization")?;

If the client hosts a Client ID Metadata Document (SEP-991), pass its URL; the SDK uses it when the server supports CIMD and falls back to dynamic registration otherwise:

oauth_state
    .start_authorization(
        AuthorizationRequest::new(MCP_REDIRECT_URI)
            .with_client_name("My MCP Client")
            .with_client_metadata_url("https://example.com/client-metadata.json"),
    )
    .await
    .context("Failed to start authorization")?;

If the client was registered with the authorization server out of band, provide the pre-registered credentials; they take priority over every other mechanism:

oauth_state
    .start_authorization(
        AuthorizationRequest::new(MCP_REDIRECT_URI)
            .with_preregistered_client("my-client-id")
            .with_client_secret("my-client-secret"),
    )
    .await
    .context("Failed to start authorization")?;

4. Get authorization url and handle callback

// get authorization URL and guide user to open it
let auth_url = oauth_state.get_authorization_url().await?;
println!("Please open the following URL in your browser for authorization:\n{}", auth_url);

// handle callback - in real applications, this is typically done in a callback server
let auth_code = "Authorization code (`code` param) obtained from browser after user authorization";
let csrf_token = "CSRF token (`state` param) obtained from browser after user authorization";
oauth_state.handle_callback(auth_code, csrf_token).await?;

5. Use Authorized Streamable HTTP Transport and create client

let am = oauth_state
    .into_authorization_manager()
    .ok_or_else(|| anyhow::anyhow!("Failed to get authorization manager"))?;
let client = AuthClient::new(reqwest::Client::default(), am);
let transport = StreamableHttpClientTransport::with_client(
    client,
    StreamableHttpClientTransportConfig::with_uri(MCP_SERVER_URL),
);

// create client and connect to MCP server
let client_service = InitializeRequestParams::default();
let client = client_service.serve(transport).await?;

If initialization reports that authorization is required, return to the application's authorization flow:

let client = match client_service.serve(transport).await {
    Ok(client) => client,
    Err(error) if error.is_authorization_required() => {
        // Prompt the user and start the application's authorization flow again.
        return Err(error.into());
    }
    Err(error) => return Err(error.into()),
};

The predicate covers both missing or expired local OAuth authorization and an HTTP 401 challenge from the MCP server. Other failures, including transient token-refresh errors and insufficient scope, return false. The original error is preserved for logging or more detailed handling; the SDK does not start an authorization flow automatically.

6. Handle scope upgrades

If a server returns 403 with insufficient_scope, you can request a scope upgrade. The SDK computes the union of current and required scopes and transitions back to the session state for re-authorization.

match oauth_state.request_scope_upgrade("admin:write", MCP_REDIRECT_URI).await {
    Ok(auth_url) => {
        // open auth_url in browser, handle callback as before
        println!("Re-authorize at: {}", auth_url);
    }
    Err(e) => {
        eprintln!("Scope upgrade failed: {}", e);
    }
}

Enterprise-managed authorization (EMA/XAA)

The example requires the rmcp features auth-enterprise-managed, client, reqwest (TLS), and transport-streamable-http-client-reqwest, plus oauth2 version 5. Call the async function from a Tokio runtime.

The exchange profile has these requirements and limits:

• Each authorization server has its own approved client registration and explicit EmaClientAuthentication: None, ClientSecretBasic, ClientSecretPost, or JwtAssertion. The SDK does not select methods from metadata or fall back to a different method after a failure.
• Input is an enterprise IdP refresh token. The requested MCP resource must match the ID-JAG's sole resource claim; scope may be omitted or narrowed.
• RAR (authorization_details) and DPoP are not supported. Nonempty authorization details are rejected at both exchange stages.
• Redemption consumes the SDK's ID-JAG handle and does not retry automatically. This is an SDK safety choice, not a protocol requirement that ID-JAGs be single-use.

The ID-JAG draft recommends confidential clients. The example below uses confidential clients with client_secret_basic at both servers. Use None only where that server permits a public client registration. Discovery, server approval, SSO, credential storage, and reauthentication remain the application's responsibility. Client-side ID-JAG checks validate structure and bindings, not signatures; the resource authorization server verifies signatures.

use oauth2::{ClientSecret, RefreshToken};
use rmcp::{
    ServiceExt,
    model::ClientInfo,
    transport::{
        StreamableHttpClientTransport,
        auth::{
            default_oauth_http_client,
            enterprise::{EmaAuthorizationServer, EmaClientAuthentication, EmaExchangeRequest},
        },
        streamable_http_client::StreamableHttpClientTransportConfig,
    },
};

async fn connect(
    refresh: &RefreshToken,
    idp_client_secret: ClientSecret,
    resource_client_secret: ClientSecret,
) -> Result<(), Box<dyn std::error::Error>> {
    let resource = "https://mcp.example/mcp";
    let http = default_oauth_http_client()?;
    let idp = EmaAuthorizationServer::new(
        "https://idp.example", "https://idp.example/token", "idp-client",
    )
    .with_client_authentication(EmaClientAuthentication::ClientSecretBasic(idp_client_secret));
    let resource_as = EmaAuthorizationServer::new(
        "https://as.example", "https://as.example/token", "mcp-client",
    )
    .with_client_authentication(EmaClientAuthentication::ClientSecretBasic(resource_client_secret));
    let token = EmaExchangeRequest::new(idp, resource_as, resource, refresh)
        .with_scopes(["files.read"])
        .exchange(&http, &http)
        .await?;
    let transport = StreamableHttpClientTransport::from_config(
        StreamableHttpClientTransportConfig::with_uri(resource)
            .auth_header(token.access_token.secret()),
    );
    let client = ClientInfo::default().serve(transport).await?;
    client.list_tools(Default::default()).await?;
    client.cancel().await?;
    Ok(())
}

auth_header takes the token without a Bearer prefix. Use it only with the approved resource and never log it. This transport uses a fixed token; obtain a new token and reconnect when it expires or is rejected.

For a registration using JWT client authentication, implement EmaClientAssertionProvider with your application's signer and configure JwtAssertion on that server. The provider example below also uses async-trait version 0.1; AppSigner represents your application's existing signing service.

use std::sync::Arc;
use rmcp::transport::auth::enterprise::{
    EmaAuthorizationServer, EmaClientAssertion, EmaClientAssertionProvider,
    EmaClientAuthentication,
};

struct AppAssertionProvider {
    signer: AppSigner,
}

#[async_trait::async_trait]
impl EmaClientAssertionProvider for AppAssertionProvider {
    async fn create_assertion(
        &self,
        server: &EmaAuthorizationServer,
    ) -> Result<EmaClientAssertion, Box<dyn std::error::Error + Send + Sync>> {
        // Sign a new assertion for this registration and approved server.
        let jwt = self.signer.sign_client_assertion(
            &server.client_id, &server.issuer, &server.token_endpoint,
        ).await?;
        Ok(EmaClientAssertion::new(jwt))
    }
}

let resource_as = EmaAuthorizationServer::new(
    "https://as.example", "https://as.example/token", "mcp-client",
)
.with_client_authentication(EmaClientAuthentication::JwtAssertion(Arc::new(
    AppAssertionProvider { signer },
)));

The SDK calls the provider before each token request, including delayed EmaIdJag::exchange redemption. Sign a fresh, short-lived assertion with a unique jti, the registered client ID in iss and sub, and the server's approved audience in aud. Your signer owns the keys and algorithm; the client assertion is separate from the ID-JAG grant. Signing and HTTP share a 30-second deadline.

The factory honors per-request redirect policy with the SDK's default reqwest settings. For custom proxy, CA, or remote-execution policy, implement OAuthHttpClient; use separate adapters for the IdP and resource AS when their network policies differ.

Complete Examples

• Authorization Code client: examples/clients/src/auth/oauth_client.rs
• Client Credentials client: examples/clients/src/auth/client_credentials.rs
• Server: examples/servers/src/complex_auth_streamhttp.rs

Running the Examples

# Run the OAuth server
cargo run -p mcp-server-examples --example servers_complex_auth_streamhttp

# Run the OAuth client (in another terminal)
cargo run -p mcp-client-examples --example clients_oauth_client

# Run the Client Credentials client
cargo run -p mcp-client-examples --example clients_client_credentials -- \
  <server_url> <client_id> <client_secret>

Authorization Flow Description

1. Resource Metadata Discovery: Client probes the server and extracts WWW-Authenticate parameters including resource_metadata URL and scope
2. Protected Resource Metadata: Client fetches resource server metadata (RFC 9728) to find authorization server(s) and supported scopes
3. AS Metadata Discovery: Client discovers authorization server metadata via RFC 8414 and OpenID Connect well-known endpoints
4. Client Registration: If supported, client dynamically registers itself (or uses URL-based Client ID via SEP-991)
5. Scope Selection: SDK picks scopes from WWW-Authenticate > PRM > AS metadata > caller defaults
6. Authorization Request: Build authorization URL with PKCE (S256) and RFC 8707 resource parameter
7. Authorization Code Exchange: After user authorization, exchange code for access token (with resource parameter)
8. Token Usage: Use access token for API calls via AuthClient or AuthorizedHttpClient
9. Token Refresh: Automatically use refresh token to get new access token when current one expires; previously granted scopes are forwarded in the refresh request so providers that require them (e.g. Azure AD v2) work correctly
10. Scope Upgrade: On 403 insufficient_scope, compute scope union and re-authorize with upgraded scopes

Security Considerations

• PKCE S256 always enforced: never falls back to plain or no challenge. OAuth 2.1 mandates S256 as Mandatory To Implement for servers.
• RFC 8707 resource binding: authorization and token requests include the resource parameter to bind tokens to the protected resource
• Redirect policy is explicit for custom OAuth clients: discovery and registration requests use OAuthHttpRedirectPolicy::Follow, while token requests use OAuthHttpRedirectPolicy::Stop so custom implementations can avoid forwarding credentials to redirected endpoints
• All tokens are securely stored in memory (custom credential stores supported)
• Automatic token refresh reduces user intervention
• Server metadata validation warns on non-compliant configurations but proceeds where relatively safe

Troubleshooting

If you encounter authorization issues, check the following:

1. Ensure server supports OAuth 2.1 authorization
2. Verify callback URI matches server's allowed redirect URIs
3. Check network connection and firewall settings
4. Verify server supports metadata discovery or dynamic client registration
5. If PKCE fails, the server may not support S256 (non-compliant with OAuth 2.1)
6. If OAuth requests need custom proxy, TLS, or connector settings, pass a configured reqwest client to OAuthState::new
7. If OAuth requests must run through a non-reqwest environment, implement OAuthHttpClient and use OAuthState::new_with_oauth_http_client
8. Check tracing logs at debug level for detailed discovery and validation info

References

• MCP Authorization Specification
• OAuth 2.1 Specification Draft
• RFC 8414: OAuth 2.0 Authorization Server Metadata
• RFC 7591: OAuth 2.0 Dynamic Client Registration Protocol
• RFC 8707: Resource Indicators for OAuth 2.0
• RFC 9728: OAuth 2.0 Protected Resource Metadata
• RFC 7636: Proof Key for Code Exchange (PKCE)
• RFC 6749 §6: Refreshing an Access Token