Looks like Python.
Compiles like Rust.

Frontier coding models score highest on Python (HumanEval, LiveCodeBench, SWE-bench Verified). lpy looks like Python so they can write it, and rustc compiles it. You get LLVM, crates.io, and about 31% less output than the same programs in Rust.

.lpy rustc LLVM native

Install

The toolchain is the lpy CLI plus rustc. This page serves it:

$ curl -sSf HOST/install.sh | sh -s -- --new ./myapp
$ lpy doctor
$ lpy check ./myapp/main.lpy
$ lpy build ./myapp/main.lpy -o ./myapp/app --opt size

That writes lpy.toml, main.lpy, AGENTS.md, and lpy.json. rustc (and cargo if you add crates) still need to be on PATH — rustup, or nix shell nixpkgs#rustc nixpkgs#cargo.

What you actually spend

At $15 per million output tokens, ten agents writing 8k tokens of Rust twenty times a task, forty tasks a day, is about $21k/month. The same work in lpy is about $15k. Edit the fields if your usage is different.


            

            

          

        

      


      

        
Playground

        
Same routes an agent uses. Check = rustc. Expand = generated Rust. Run needs def main().

        

          
          
          
          
        

        

          
          
Output from check / expand / run / tokens shows up here.

        

      


      

        
Give this to your agent

        
No training. Tell it to write .lpy, GET /llms.txt, and POST source at HOST. Models are already good at Python. rustc does the checking.

        

        
Same CLI as Install. After curl … | sh -s -- --new ./myapp, point the agent at that folder.

        

          
$ curl -sS HOST/llms.txt

          
$ curl -sS -X POST HOST/api/check --data-binary @hello.lpy

          
$ curl -sS -X POST HOST/api/run --data-binary @hello.lpy

        

      


      

        
What you get

        
It uses Python syntax so models can write it. rustc builds a native binary. Libraries come from crates.io. Compile with lpy build.