diff --git a/bin/pudox-run/src/main.rs b/bin/pudox-run/src/main.rs
index 4cb6566..98083e5 100644
--- a/bin/pudox-run/src/main.rs
+++ b/bin/pudox-run/src/main.rs
@@ -8,9 +8,17 @@ use async_openai::types::chat::{
 };
 use async_openai::Client;
 
-const PUDOXD_TASK_URL: &str = "http://localhost:7790/task";
-const PUDOXD_RESULT_URL: &str = "http://localhost:7790/task/result";
-const PUDOXD_STREAM_URL: &str = "http://localhost:7790/task/stream";
+macro_rules! pudoxd_url {
+    ($path:literal) => {
+        concat!("http://localhost:7790", $path)
+    };
+}
+
+const PUDOXD_OPENAI_URL: &str = pudoxd_url!("/v1");
+const PUDOXD_TASK_URL: &str = pudoxd_url!("/task");
+const PUDOXD_RESULT_URL: &str = pudoxd_url!("/task/result");
+const PUDOXD_STREAM_URL: &str = pudoxd_url!("/task/stream");
+const PUDOXD_INFERENCE_URL: &str = pudoxd_url!("/v1/chat/completions");
 
 #[tokio::main]
 async fn main() -> Result<(), Box<dyn std::error::Error>> {
@@ -123,7 +131,7 @@ async fn process_task(
 ) -> Result<TaskResult, Box<dyn std::error::Error>> {
     let client = Client::with_config(
         OpenAIConfig::new()
-            .with_api_base("http://ai")
+            .with_api_base(PUDOXD_OPENAI_URL)
             .with_api_key("none"),
     );
 
@@ -293,7 +301,7 @@ async fn summarize_session(
     });
 
     let resp: serde_json::Value = reqwest::Client::new()
-        .post("http://ai/v1/chat/completions")
+        .post(PUDOXD_INFERENCE_URL)
         .json(&body)
         .send()
         .await?
@@ -342,7 +350,7 @@ async fn generate_pull_request_info(
     for attempt in 1..=3u32 {
         tracing::info!("generate_pull_request_info: attempt {}/3", attempt);
         let resp: serde_json::Value = reqwest::Client::new()
-            .post("http://ai/v1/chat/completions")
+            .post(PUDOXD_INFERENCE_URL)
             .json(&body)
             .send()
             .await?
diff --git a/bin/pudoxd/Cargo.toml b/bin/pudoxd/Cargo.toml
index 2db1869..25fa47e 100644
--- a/bin/pudoxd/Cargo.toml
+++ b/bin/pudoxd/Cargo.toml
@@ -12,7 +12,7 @@ axum = "0.7"
 tokio = { version = "1", features = ["full"] }
 tracing = "0.1.44"
 tracing-subscriber = "0.3.23"
-reqwest = { version = "0.12", features = ["json"] }
+reqwest = { version = "0.12", features = ["json", "stream"] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 git2 = { version = "0.20", default-features = false, features = ["vendored-libgit2", "vendored-openssl", "https"] }
diff --git a/bin/pudoxd/src/main.rs b/bin/pudoxd/src/main.rs
index 3d3fc9e..769d70e 100644
--- a/bin/pudoxd/src/main.rs
+++ b/bin/pudoxd/src/main.rs
@@ -1,6 +1,7 @@
 use axum::{
+    body::Body,
     extract::State,
-    http::StatusCode,
+    http::{Request, Response, StatusCode},
     response::{
         sse::{Event, KeepAlive, Sse},
         IntoResponse,
@@ -25,6 +26,7 @@ use crate::session::{run_session, Session, SessionMessage};
 struct App {
     linear: Linear,
     sessions: Arc<RwLock<HashMap<String, Session>>>,
+    http: reqwest::Client,
 }
 
 #[tokio::main]
@@ -36,6 +38,7 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
     let app = App {
         linear: Linear::new(),
         sessions: Arc::new(RwLock::new(HashMap::new())),
+        http: reqwest::Client::new(),
     };
 
     let router = Router::new()
@@ -43,6 +46,7 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
         .route("/task", get(task_handler))
         .route("/task/result", post(task_result_handler))
         .route("/task/stream", get(stream_handler))
+        .route("/v1/chat/completions", post(chat_completions_proxy))
         .with_state(app);
 
     let addr = SocketAddr::from(([0, 0, 0, 0], 7790));
@@ -344,6 +348,57 @@ async fn handle_task_result(
     }
 }
 
+async fn chat_completions_proxy(State(app): State<App>, req: Request<Body>) -> impl IntoResponse {
+    let content_type = req
+        .headers()
+        .get("content-type")
+        .and_then(|v| v.to_str().ok())
+        .unwrap_or("application/json")
+        .to_string();
+
+    let body_bytes = match axum::body::to_bytes(req.into_body(), usize::MAX).await {
+        Ok(b) => b,
+        Err(_) => {
+            return Response::builder()
+                .status(StatusCode::BAD_GATEWAY)
+                .body(Body::empty())
+                .unwrap();
+        }
+    };
+
+    let upstream = match app
+        .http
+        .post("http://ai/v1/chat/completions")
+        .header("content-type", content_type)
+        .body(body_bytes)
+        .send()
+        .await
+    {
+        Ok(r) => r,
+        Err(e) => {
+            tracing::error!("Upstream inference request failed: {}", e);
+            return Response::builder()
+                .status(StatusCode::BAD_GATEWAY)
+                .body(Body::empty())
+                .unwrap();
+        }
+    };
+
+    let status = upstream.status();
+    let resp_content_type = upstream
+        .headers()
+        .get("content-type")
+        .and_then(|v| v.to_str().ok())
+        .unwrap_or("application/json")
+        .to_string();
+
+    Response::builder()
+        .status(status)
+        .header("content-type", resp_content_type)
+        .body(Body::from_stream(upstream.bytes_stream()))
+        .unwrap()
+}
+
 async fn stream_handler(
     State(app): State<App>,
     headers: axum::http::HeaderMap,