winx_code_agent/
http_server.rs

1//! Optional remote MCP transport over Streamable HTTP.
2//!
3//! The endpoint exposes shell and filesystem capabilities, so remote mode is
4//! deliberately fail-closed: loopback bind by default, strong file-backed or
5//! explicit bearer credentials, bounded request concurrency/rate, and isolated
6//! daemon namespaces for independently authenticated principals.
7
8#![allow(clippy::doc_markdown)]
9
10use std::collections::{HashMap, HashSet};
11use std::net::{IpAddr, SocketAddr};
12use std::sync::Arc;
13use std::time::{Duration, Instant};
14
15use axum::{
16    extract::{ConnectInfo, Request, State},
17    http::{header, HeaderValue, StatusCode},
18    middleware::{self, Next},
19    response::{IntoResponse, Response},
20    Router,
21};
22use rmcp::transport::streamable_http_server::{
23    session::local::LocalSessionManager, StreamableHttpServerConfig, StreamableHttpService,
24};
25use tokio::sync::{Mutex, Semaphore};
26use tower_http::limit::RequestBodyLimitLayer;
27use tower_http::timeout::TimeoutLayer;
28
29use crate::config::{HttpPrincipal, HttpSessionAffinity};
30use crate::runtime::{configured_shell_runtime, ShellRuntime};
31use crate::server::{SessionIsolation, WinxService};
32
33pub type BoxError = Box<dyn std::error::Error + Send + Sync>;
34
35const MAX_BODY_BYTES: usize = 64 * 1024 * 1024;
36const REQUEST_TIMEOUT: Duration = Duration::from_secs(120);
37const RATE_WINDOW: Duration = Duration::from_secs(60);
38const MAX_RATE_TRACKED_IPS: usize = 4_096;
39const INVALID_AUTH_DELAY: Duration = Duration::from_millis(100);
40pub const DEFAULT_MAX_CONCURRENCY: usize = 32;
41pub const DEFAULT_REQUESTS_PER_MINUTE: u32 = 120;
42
43#[derive(Clone, Debug)]
44pub struct HttpServerOptions {
45    pub bind: String,
46    pub principals: Vec<HttpPrincipal>,
47    pub extra_hosts: Vec<String>,
48    pub allow_query_token: bool,
49    pub allow_non_loopback: bool,
50    pub session_affinity: HttpSessionAffinity,
51    pub max_concurrency: usize,
52    pub requests_per_minute: u32,
53}
54
55impl HttpServerOptions {
56    pub fn single(
57        bind: impl Into<String>,
58        token: impl Into<String>,
59        extra_hosts: Vec<String>,
60        allow_query_token: bool,
61    ) -> Result<Self, BoxError> {
62        let token = token.into();
63        let principal = HttpPrincipal::new("default", token.trim(), false)?;
64        Ok(Self {
65            bind: bind.into(),
66            principals: vec![principal],
67            extra_hosts,
68            allow_query_token,
69            allow_non_loopback: false,
70            session_affinity: HttpSessionAffinity::Workspace,
71            max_concurrency: DEFAULT_MAX_CONCURRENCY,
72            requests_per_minute: DEFAULT_REQUESTS_PER_MINUTE,
73        })
74    }
75}
76
77pub async fn start_http_server(
78    bind: &str,
79    token: String,
80    extra_hosts: Vec<String>,
81    allow_query_token: bool,
82) -> Result<(), BoxError> {
83    let options = HttpServerOptions::single(bind, token, extra_hosts, allow_query_token)?;
84    start_http_server_with_options(options).await
85}
86
87pub async fn start_http_server_with_options(options: HttpServerOptions) -> Result<(), BoxError> {
88    let shell_runtime = configured_shell_runtime().await?;
89    start_http_server_with_runtime(options, shell_runtime).await
90}
91
92pub async fn start_http_server_with_runtime(
93    options: HttpServerOptions,
94    shell_runtime: Arc<dyn ShellRuntime>,
95) -> Result<(), BoxError> {
96    let bind = validate_options(&options)?;
97    let principal_names =
98        options.principals.iter().map(|principal| principal.name().to_string()).collect::<Vec<_>>();
99    let shared_concurrency = Arc::new(Semaphore::new(options.max_concurrency));
100    let shared_rate_limiter = Arc::new(RateLimiter::new(options.requests_per_minute));
101
102    let mut config = StreamableHttpServerConfig::default();
103    config.legacy_session_mode = true;
104    config.allowed_hosts.extend(options.extra_hosts);
105
106    let shared = WinxService::with_runtime(SessionIsolation::Strict, shell_runtime);
107    let mcp_service = StreamableHttpService::new(
108        move || Ok(shared.clone()),
109        Arc::new(LocalSessionManager::default()),
110        config,
111    );
112    let security = Arc::new(RouteSecurity {
113        principals: Arc::new(options.principals),
114        allow_query: options.allow_query_token,
115        session_affinity: options.session_affinity,
116        concurrency: shared_concurrency,
117        rate_limiter: shared_rate_limiter,
118    });
119    // The timeout sits INSIDE the security middleware: when it fires, the
120    // middleware still receives the 408 response and records the usage event.
121    // With the previous order the timed-out future was dropped before logging,
122    // so 408s were invisible in the usage log.
123    let protected_mcp = Router::new()
124        .nest_service("/mcp", mcp_service)
125        .layer(TimeoutLayer::with_status_code(StatusCode::REQUEST_TIMEOUT, REQUEST_TIMEOUT))
126        .layer(middleware::from_fn_with_state(security, enforce_route_security))
127        .layer(RequestBodyLimitLayer::new(MAX_BODY_BYTES));
128    // Keep OAuth discovery probes and every unrelated path outside bearer auth.
129    // Winx does not advertise OAuth metadata, so those routes should be ordinary
130    // 404s instead of 401s that encourage clients to start a nonexistent flow.
131    let app = Router::new().merge(protected_mcp);
132
133    let listener = tokio::net::TcpListener::bind(bind).await?;
134    tracing::warn!(
135        bind = %bind,
136        principals = ?principal_names,
137        route = "/mcp",
138        "Winx remote MCP is network-reachable; keep it behind authenticated HTTPS"
139    );
140    if options.allow_query_token {
141        tracing::warn!(
142            "query-token authentication is enabled; URLs may expose credentials in proxy logs/history"
143        );
144    }
145    axum::serve(listener, app.into_make_service_with_connect_info::<SocketAddr>()).await?;
146    Ok(())
147}
148
149fn validate_options(options: &HttpServerOptions) -> Result<SocketAddr, BoxError> {
150    let bind: SocketAddr = options
151        .bind
152        .parse()
153        .map_err(|error| format!("invalid --bind socket address {:?}: {error}", options.bind))?;
154    if !bind.ip().is_loopback() && !options.allow_non_loopback {
155        return Err(format!(
156            "refusing non-loopback HTTP bind {bind}; pass --allow-non-loopback only behind a trusted firewall/tunnel"
157        )
158        .into());
159    }
160    if options.max_concurrency == 0 {
161        return Err("max_concurrency must be at least 1".into());
162    }
163    if options.requests_per_minute == 0 {
164        return Err("requests_per_minute must be at least 1".into());
165    }
166    if options.principals.is_empty() {
167        return Err("at least one HTTP principal is required".into());
168    }
169    let mut names = HashSet::new();
170    for principal in &options.principals {
171        if !names.insert(principal.name()) {
172            return Err(format!("duplicate HTTP principal: {}", principal.name()).into());
173        }
174    }
175    Ok(bind)
176}
177
178#[derive(Clone)]
179struct RouteSecurity {
180    principals: Arc<Vec<HttpPrincipal>>,
181    allow_query: bool,
182    session_affinity: HttpSessionAffinity,
183    concurrency: Arc<Semaphore>,
184    rate_limiter: Arc<RateLimiter>,
185}
186
187async fn enforce_route_security(
188    State(security): State<Arc<RouteSecurity>>,
189    ConnectInfo(peer): ConnectInfo<SocketAddr>,
190    request: Request,
191    next: Next,
192) -> Response {
193    let path = request.uri().path();
194    if path != "/mcp" && !path.starts_with("/mcp/") {
195        return next.run(request).await;
196    }
197    let started = Instant::now();
198    if !security.rate_limiter.allow(peer.ip()).await {
199        return retry_response(StatusCode::TOO_MANY_REQUESTS, "request rate limit exceeded\n");
200    }
201    let Ok(permit) = security.concurrency.clone().try_acquire_owned() else {
202        return retry_response(StatusCode::SERVICE_UNAVAILABLE, "too many concurrent requests\n");
203    };
204
205    let Some(principal) =
206        authenticate_request(&request, &security.principals, security.allow_query)
207    else {
208        tokio::time::sleep(INVALID_AUTH_DELAY).await;
209        tracing::warn!(
210            peer = %peer,
211            path = %request.uri().path(),
212            "rejected request with missing or invalid HTTP token"
213        );
214        return (StatusCode::UNAUTHORIZED, "missing or invalid token\n").into_response();
215    };
216
217    let principal_name = principal.name().to_string();
218    let method = request.method().clone();
219    let mut request = request;
220    request.extensions_mut().insert(principal);
221    request.extensions_mut().insert(security.session_affinity);
222    let response = next.run(request).await;
223    drop(permit);
224    // One structured `winx::usage` event per authenticated request. Metadata
225    // only — the body (commands, file contents) is never logged.
226    tracing::info!(
227        target: "winx::usage",
228        event = "http_request",
229        principal = %principal_name,
230        peer = %peer,
231        method = %method,
232        status = response.status().as_u16(),
233        duration_ms = u64::try_from(started.elapsed().as_millis()).unwrap_or(u64::MAX),
234        "http request"
235    );
236    response
237}
238
239fn retry_response(status: StatusCode, body: &'static str) -> Response {
240    let mut response = (status, body).into_response();
241    response.headers_mut().insert(header::RETRY_AFTER, HeaderValue::from_static("1"));
242    response
243}
244
245#[derive(Debug)]
246struct RateWindow {
247    started_at: Instant,
248    last_seen: Instant,
249    count: u32,
250}
251
252#[derive(Debug)]
253struct RateLimiter {
254    requests_per_window: u32,
255    windows: Mutex<HashMap<IpAddr, RateWindow>>,
256}
257
258impl RateLimiter {
259    fn new(requests_per_window: u32) -> Self {
260        Self { requests_per_window, windows: Mutex::new(HashMap::new()) }
261    }
262
263    async fn allow(&self, peer: IpAddr) -> bool {
264        let now = Instant::now();
265        let mut windows = self.windows.lock().await;
266        windows.retain(|_, window| now.duration_since(window.last_seen) < RATE_WINDOW * 2);
267        if !windows.contains_key(&peer) && windows.len() >= MAX_RATE_TRACKED_IPS {
268            if let Some(oldest) =
269                windows.iter().min_by_key(|(_, window)| window.last_seen).map(|(ip, _)| *ip)
270            {
271                windows.remove(&oldest);
272            }
273        }
274        let window =
275            windows.entry(peer).or_insert(RateWindow { started_at: now, last_seen: now, count: 0 });
276        if now.duration_since(window.started_at) >= RATE_WINDOW {
277            window.started_at = now;
278            window.count = 0;
279        }
280        window.last_seen = now;
281        if window.count >= self.requests_per_window {
282            return false;
283        }
284        window.count = window.count.saturating_add(1);
285        true
286    }
287}
288
289fn authenticate_request(
290    request: &Request,
291    principals: &[HttpPrincipal],
292    allow_query: bool,
293) -> Option<HttpPrincipal> {
294    let header = bearer_token(request);
295    let query = allow_query.then(|| query_token(request)).flatten();
296    [header, query].into_iter().flatten().find_map(|presented| {
297        principals.iter().find(|principal| principal.matches_token(presented)).cloned()
298    })
299}
300
301fn bearer_token(request: &Request) -> Option<&str> {
302    let value = request.headers().get(header::AUTHORIZATION)?.to_str().ok()?;
303    let (scheme, token) = value.split_once(' ')?;
304    scheme.eq_ignore_ascii_case("bearer").then_some(token.trim())
305}
306
307fn query_token(request: &Request) -> Option<&str> {
308    request.uri().query()?.split('&').find_map(|pair| pair.strip_prefix("token="))
309}
310
311#[cfg(test)]
312mod tests {
313    #![allow(clippy::expect_used)]
314
315    use std::net::{IpAddr, Ipv4Addr};
316
317    use axum::body::Body;
318    use axum::extract::Request;
319
320    use super::{authenticate_request, validate_options, HttpServerOptions, RateLimiter};
321    use crate::config::HttpPrincipal;
322
323    fn request(uri: &str, authorization: Option<&str>) -> Request {
324        let mut builder = Request::builder().uri(uri);
325        if let Some(authorization) = authorization {
326            builder = builder.header(axum::http::header::AUTHORIZATION, authorization);
327        }
328        builder.body(Body::empty()).expect("test request")
329    }
330
331    fn principal(name: &str, token_byte: char) -> HttpPrincipal {
332        let token = token_byte.to_string().repeat(32);
333        HttpPrincipal::new(name, &token, false).expect("valid principal")
334    }
335
336    fn strong_options(bind: &str) -> HttpServerOptions {
337        HttpServerOptions {
338            bind: bind.to_string(),
339            principals: vec![principal("default", 'x')],
340            extra_hosts: Vec::new(),
341            allow_query_token: false,
342            allow_non_loopback: false,
343            session_affinity: crate::config::HttpSessionAffinity::Workspace,
344            max_concurrency: super::DEFAULT_MAX_CONCURRENCY,
345            requests_per_minute: super::DEFAULT_REQUESTS_PER_MINUTE,
346        }
347    }
348
349    #[test]
350    fn bearer_and_opt_in_query_tokens_select_the_matching_principal() {
351        let principals = vec![principal("chatgpt", 'a'), principal("automation", 'b')];
352        let header = request("/mcp", Some(&format!("Bearer {}", "a".repeat(32))));
353        let matched = authenticate_request(&header, &principals, false).expect("header principal");
354        assert_eq!(matched.name(), "chatgpt");
355
356        let query_uri = format!("/mcp?token={}", "b".repeat(32));
357        assert!(authenticate_request(&request(&query_uri, None), &principals, false).is_none());
358        let matched = authenticate_request(&request(&query_uri, None), &principals, true)
359            .expect("query principal");
360        assert_eq!(matched.name(), "automation");
361    }
362
363    #[test]
364    fn non_loopback_binding_requires_explicit_override() {
365        assert!(validate_options(&strong_options("127.0.0.1:8000")).is_ok());
366        assert!(validate_options(&strong_options("0.0.0.0:8000")).is_err());
367
368        let mut public = strong_options("0.0.0.0:8000");
369        public.allow_non_loopback = true;
370        assert!(validate_options(&public).is_ok());
371    }
372
373    #[tokio::test]
374    async fn rate_limiter_rejects_excess_requests() {
375        let limiter = RateLimiter::new(2);
376        let peer = IpAddr::V4(Ipv4Addr::LOCALHOST);
377        assert!(limiter.allow(peer).await);
378        assert!(limiter.allow(peer).await);
379        assert!(!limiter.allow(peer).await);
380    }
381
382    #[test]
383    fn duplicate_principal_names_are_rejected() {
384        let mut options = strong_options("127.0.0.1:8000");
385        options.principals = vec![principal("same", 'a'), principal("same", 'b')];
386        assert!(validate_options(&options).is_err());
387    }
388}
winx_code_agent/http_server.rs

winx_code_agent/
http_server.rs
