* fix(auth): preserve OAuth discovery transport errors * refactor!: expose domain-specific boxed OAuth HTTP errors Let custom OAuth HTTP clients preserve native error types and source chains while keeping the type-erased boundary explicit in the public API. BREAKING CHANGE: OAuthHttpClientError is now a boxed error alias; custom clients should return native errors with .into() or box them directly. --------- Co-authored-by: Theodore Ni <3806110+tjni@users.noreply.github.com>
15 KiB
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
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:
- Pre-registered client information (
with_preregistered_client), when the client already holds aclient_idissued out of band - Client ID Metadata Documents (SEP-991,
with_client_metadata_url), when the authorization server advertisesclient_id_metadata_document_supported - 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 = ClientInfo::default();
let client = client_service.serve(transport).await?;
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);
}
}
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
- Resource Metadata Discovery: Client probes the server and extracts
WWW-Authenticateparameters includingresource_metadataURL andscope - Protected Resource Metadata: Client fetches resource server metadata (RFC 9728) to find authorization server(s) and supported scopes
- AS Metadata Discovery: Client discovers authorization server metadata via RFC 8414 and OpenID Connect well-known endpoints
- Client Registration: If supported, client dynamically registers itself (or uses URL-based Client ID via SEP-991)
- Scope Selection: SDK picks scopes from WWW-Authenticate > PRM > AS metadata > caller defaults
- Authorization Request: Build authorization URL with PKCE (S256) and RFC 8707 resource parameter
- Authorization Code Exchange: After user authorization, exchange code for access token (with resource parameter)
- Token Usage: Use access token for API calls via
AuthClientorAuthorizedHttpClient - 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
- 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
plainor no challenge. OAuth 2.1 mandates S256 as Mandatory To Implement for servers. - RFC 8707 resource binding: authorization and token requests include the
resourceparameter 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 useOAuthHttpRedirectPolicy::Stopso 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:
- Ensure server supports OAuth 2.1 authorization
- Verify callback URI matches server's allowed redirect URIs
- Check network connection and firewall settings
- Verify server supports metadata discovery or dynamic client registration
- If PKCE fails, the server may not support S256 (non-compliant with OAuth 2.1)
- If OAuth requests need custom proxy, TLS, or connector settings, pass a configured reqwest client to
OAuthState::new - If OAuth requests must run through a non-reqwest environment, implement
OAuthHttpClientand useOAuthState::new_with_oauth_http_client - Check
tracinglogs 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