Skip to content
StatusQuickstartDocsArchitectureDesign Run locally
What is real

Serving HTTP and AI from swift-os

A pure-Swift in-kernel TCP/IP stack, a real HTTP server, and a TinyStories inference daemon — what is shipping, and what is still ahead.


swift-os does more than print to a console. Let us be precise about what is live and what is ahead.

What is live today

The network stack is a pure-Swift, sans-IO core — Ethernet/ARP, IPv4, ICMP, UDP, TCP, and DNS — with no per-packet heap allocation, driving a modern virtio-net device. On top of it:

  • /bin/httpd is a real concurrent, poll()-driven static file server with MIME types and directory listings, on :8080.
  • /bin/llmd serves a TinyStories transformer over HTTP: GET /health, POST /completion, GET /metrics. The transformer and tokenizer are written in Embedded Swift and compiled into a static EL0 binary.

Both run unprivileged, holding only a net:bind capability.

"Hello from the kernel" is where most hobby OSes stop. Answering a real HTTP request end-to-end — NIC → in-kernel TCP → an unprivileged Swift handler — is a much harder, much more convincing demonstration.

What is honest about it

The current llmd is a TinyStories proof of the application & AI-hosting profile — not a general ONNX/GGUF/PyTorch or GPU runtime. TLS 1.3 has a working client path (ChaCha20-Poly1305), but production certificate verification is incomplete. Node.js and JVM hosting are recorded Phase-2 design, not implementation.

Why it matters

The flagship profile is application and AI hosting. A native HTTP server and a native inference daemon, both in Embedded Swift on a capability-secured base, are the concrete first proof that the profile is more than a slogan.