diff --git a/rust/Cargo.toml b/rust/Cargo.toml
index af8dd531d..943cd515b 100644
--- a/rust/Cargo.toml
+++ b/rust/Cargo.toml
@@ -1,6 +1,10 @@
 [workspace]
 resolver = "3"
-members = ["sglang-grpc", "sglang-mm"]
+members = [
+    "sglang-grpc",
+    "sglang-mm",
+    "sglang-server"
+]
 
 [workspace.package]
 version = "0.1.0"
@@ -12,11 +16,16 @@ license = "Apache-2.0"
 # where needed).
 [workspace.dependencies]
 async-stream = "0.3"
+bytes = "1"
+futures = "0.3"
 pyo3 = { version = "0.29.0", features = ["extension-module"] }
+serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+thiserror = "2"
 tokio = { version = "1", features = ["full"] }
 tokio-stream = { version = "0.1", features = ["net"] }
 tracing = "0.1"
+tracing-appender = "0.2"
 tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 uuid = { version = "1", features = ["v4"] }
 
diff --git a/rust/sglang-server/Cargo.toml b/rust/sglang-server/Cargo.toml
new file mode 100644
index 000000000..71f6023d1
--- /dev/null
+++ b/rust/sglang-server/Cargo.toml
@@ -0,0 +1,40 @@
+[package]
+name = "sglang-server"
+description = "Sglang server in rust"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+
+# Consumed by python/setup.py: registers this crate as a PyO3 extension module.
+[package.metadata.sglang]
+python-module = "sglang.srt.server._core"
+
+[lib]
+name = "sglang_server"
+# cdylib: the pyo3 module imported by the (embedded) Python scheduler process.
+# rlib:   so optional standalone bins / integration tests can link the core.
+crate-type = ["cdylib", "rlib"]
+
+[dependencies]
+async-stream = { workspace = true }
+bytes = { workspace = true }
+futures = { workspace = true }
+pyo3 = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+thiserror = { workspace = true }
+tokio = { workspace = true }
+tracing = { workspace = true }
+tracing-subscriber = { workspace = true }
+tracing-appender = { workspace = true }
+uuid = { workspace = true }
+
+axum = { version = "0.8.9", features = ["json", "tokio"] }
+core_affinity = "0.8"
+# the dynamo-tokenizers is deps on hf-hub, should bump version together
+dynamo-tokenizers = "1.5.3"
+flume = "0.12.0"
+hf-hub = { version = "0.4", default-features = false }
+regex-syntax = "0.8"
+rmp-serde = "1"
+rmpv = { version = "1", features = ["with-serde"] }
diff --git a/rust/sglang-server/pyproject.toml b/rust/sglang-server/pyproject.toml
new file mode 100644
index 000000000..1d2c1c424
--- /dev/null
+++ b/rust/sglang-server/pyproject.toml
@@ -0,0 +1,16 @@
+[build-system]
+requires = ["maturin>=1.13,<2.0"]
+build-backend = "maturin"
+
+[project]
+name = "sglang-server"
+requires-python = ">=3.10"
+classifiers = [
+    "Programming Language :: Rust",
+    "Programming Language :: Python :: Implementation :: CPython",
+    "Programming Language :: Python :: Implementation :: PyPy",
+]
+dynamic = ["version"]
+
+[tool.maturin]
+module-name = "_core"
diff --git a/rust/sglang-server/src/lib.rs b/rust/sglang-server/src/lib.rs
new file mode 100644
index 000000000..ab23f963c
--- /dev/null
+++ b/rust/sglang-server/src/lib.rs
@@ -0,0 +1,19 @@
+//! sglang-server: a multi-threaded Rust frontend (API server → TokenizerManager
+//! → Tokenizer/Detokenizer) embedded in the Python scheduler process.
+//!
+//! Pipeline stages 1–5 are pure Rust and never touch a `PyObject`, so they run
+//! concurrently with the Python scheduler without contending for the GIL. The
+//! only GIL crossings are the boundary methods on [`Server`]:
+//!   * `recv_requests` — Python scheduler thread drains the ingress ring.
+//!   * `push_batch`    — Python scheduler thread pushes one output batch.
+//!   * `push_result`   — Python scheduler thread pushes one control result.
+//!
+//! All are non-blocking, so the GIL is never held across a wait.
+#![allow(dead_code)] // TODO: remove when the consumer PR lands
+
+use pyo3::prelude::*;
+
+#[pymodule]
+fn _core(_m: &Bound<'_, PyModule>) -> PyResult<()> {
+    Ok(())
+}