smlt

/smɛlt/ — to form ore into metal.

Why

We stopped reading agent-generated code. PRs get approved with a mass thumbs-up, diffs scroll past unreviewed, the code "works" and nobody questions the stack underneath. If the human is no longer the bottleneck — no longer the one checking every line — then why keep generating bloated, interpreted, runtime-dependent code?

smlt takes the logical next step: let the agent write code as close to the metal as it gets. The LLM acts as a compiler frontend, emitting portable LLVM IR from natural language. The result is small, fast, dependency-free native binaries. No runtime. No garbage collector. No node_modules. Just machine code.

Install

curl -fsSL https://github.com/smltdev/smlt/releases/latest/download/smlt-$(uname -s | tr A-Z a-z)-$(uname -m | sed 's/aarch64/arm64/') -o /usr/local/bin/smlt
chmod +x /usr/local/bin/smlt

Then fetch the core toolchain (modules, platform libs):

smlt fetch core

This clones the core repo into $SMLT_HOME (defaults to ~/.local/share/smlt). You can override it:

export SMLT_HOME=/path/to/smlt
smlt fetch core

Requirements: clang and llvm-link.

# macOS
xcode-select --install
brew install llvm

# Ubuntu / Debian
apt install clang llvm

Install the Claude Code skill

The skill teaches Claude how to emit LLVM IR, use the module system, and build native binaries.

For projects inside the core repo — it works automatically. Claude picks it up from the project directory.

For any other project — run from your project directory:

smlt install-skill

This downloads the latest skill into .claude/skills/smlt/. For Codex or other agents that use .agents/:

smlt install-skill --agents

Usage

smlt build app.ll output --auto          # build with auto-detected modules
smlt build app.ll output --use buf       # explicit module selection
smlt cross app.ll myapp --auto           # cross-compile for all platforms
smlt install app.ll                      # build and install to $SMLT_HOME/bin/
smlt run app.ll -- args                  # build to temp, execute, clean up
smlt fetch <module>                      # fetch module from registry
smlt install-skill                       # install Claude Code skill
smlt doctor                              # check build environment

Architecture

natural language
      |
  Claude (compiler frontend)
      |
  app.ll              portable IR — no target triple, no OS constants
      |
      + modules/       pre-built, audited IR modules (buf, http, str, ...)
      + platform/      per-target platform abstraction
      |
  smlt build  →  native binary

App IR declares module symbols. The build system resolves dependencies, links the correct platform module, and compiles to a native binary. LTO inlines everything — zero runtime overhead.

Modules

Pre-built modules cover common infrastructure so the LLM generates only app-specific logic:

Module	Purpose
buf	Bounds-checked buffer writes, JSON escaping, IP formatting
cli	Argument parsing, help text formatting
ev	Async event loop (io_uring on Linux, kqueue on macOS)
fs	File I/O: read, write, append, exists, mkdir
hex	Hex encoding/decoding
http	Async HTTP/1.1 server (callback-based)
httpc	HTTP/1.1 client with TLS/HTTPS support
proc	Fork/exec, stdout capture
sha256	SHA-256 hashing (incremental or one-shot)
sort	In-place insertion sort
str	String comparison, search, integer conversion
tar	Create and iterate tar archives in memory
tls	OpenSSL/LibreSSL wrapper for client and server TLS

Modules are portable IR — no target triple, no OS constants. Each has a manifest.yaml with symbols and dependencies. See modules/buf/ as the reference implementation.

Platform abstraction

All cross-platform differences are hidden behind platform_* calls:

	macOS	Linux
SOL_SOCKET	65535	1
SO_REUSEADDR	4	2
sockaddr_in byte 0	sin_len (u8)	sin_family (u16)
addrinfo.ai_addr	offset 32	offset 24

Supported platforms: darwin-arm64, darwin-x86_64, linux-x86_64, linux-arm64.

Files

platform/                    Per-target platform modules
modules/                     Pre-built IR module library
apps/smlt/                   smlt CLI tool (built with smlt)
apps/fetch/                  curl-like HTTP/HTTPS client
apps/registry/               Module registry server
examples/                    Example applications
.claude/skills/smlt/         Claude Code skill for IR codegen