A programming language for AI agents

Code an agent writes. Limits you can read.

Every Aleph file states what it may touch, carries its own tests and answers mistakes in JSON a model can act on. You approve a program by reading three lines, not three hundred.

Read the 2,000-token specView the compiler

Rust compiler, C backend, no runtime dependencies. Open source on GitHub.

words.aph8 lines · 1 capability · 1 test
use fs.read "data"
fn words s:str -> i64
 ex words "a b c" = 3
 len split trim s " "
fn main -> res unit str
 let text try read "data/in.txt"
 do print show words text
 ok nil
$ aleph explain words.aphcapabilities: fs.read "data"
words: pure · 1 example
main: effectful · calls words
$ aleph check (without the use line)"code":"C001", "node":"n7",
"msg":"missing capability",
"expected":"fs.read"

Read it in order, or not

Five ways into Aleph.

Epilogue

What it sees is simultaneous. What you read is successive.

That was the narrator’s despair: language unrolls in a line what the eye took in at once. Aleph splits the work. The agent holds the whole file; you read three lines of aleph explain. To begin, give your agent the spec.

One file, generated from the compiler's own tables so it never drifts: keywords, builtins with their types and capabilities, diagnostics and ten examples that pass check.

$ cargo install --git https://github.com/qursor-it/aleph alephc
$ aleph llms > context.txt
$ aleph check words.aph && aleph build words.aph -o words
$ ./words
3