LaC (LLM as Code) - reference engine. Declarative behavior for LLM apps: law in files, perimeter in code.
Project description
LaC - LLM as Code
LaC is a protocol for building LLM applications the way IaC builds infrastructure: behavior is declared in versioned artifacts, not improvised in prompts. An application is a folder of plain files - a compose declaration, law files with explicit authority levels, and a commands module. The engine is generic and knows nothing about any particular application.
This repository holds the LaC specification (SPEC.md) and the reference engine. The engine is intentionally small: a compose loader with a strict parser, a context assembler that loads law in trust order (L1 then L2 then L3), a REPL, and an agentic loop where free-form language and canonical !commands are two roads into the same code.
Status: pre-release. The spec (SPEC.md, methodology v0.2) is the canon; this engine tracks it through milestones M0-M3 (loader, retrieval commands, write perimeter, drift metrics). APIs and the compose schema will change.
Core ideas
- Levels are roles. L1 (admin) and L2 (dev) files carry behavior and load always; L3 (user data) is content, never instructions.
- Two roads, one code. Canonical
!commandsbypass the model entirely (zero tokens); free-form requests reach the same functions through tool calling. The model interprets and narrates; code executes. - Perimeter in code, not in prose. Effects on disk are the engine's job. A wrong action must be survivable.
Quick start
pipx install git+https://github.com/diranix/lac
cd your-app
lac
(PyPI release is planned; until then the package installs straight from this repo.)
An application folder looks like:
your-app/
.lac/
llm_compose.yaml # the anchor: all paths resolve relative to the app root
law/ # L1/L2 law files
souls/ # personas (L2)
commands_<app>.py # the app's command module (TOOLS + COMMANDS registry)
<memory folders> # L3 content
The engine finds .lac/llm_compose.yaml in the current directory (or in the directory passed as the first argument) and resolves every path in the compose against that root. No absolute paths anywhere.
Providers: Anthropic (raw HTTP, prompt caching, ANTHROPIC_API_KEY from env only) and Ollama for local models. The vendor surface lives in one function, adapter.send().
Trust model: the commands module is application code (L2) and the engine executes it deliberately - running an app means trusting its dev layer, the same way installing any package means trusting its authors. The levels protect L1/L2 from the model and from L3 content, not the machine from the app you chose to run.
First application
Grimoire - a personal memory system and the first LaC application. Its earlier implementation runs the same protocol on Claude Code, with the perimeter enforced by harness permissions instead of engine code: one law, two hosts.
License
Apache-2.0. The protocol is free; applications built on it license themselves independently.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file lac_engine-0.1.0.tar.gz.
File metadata
- Download URL: lac_engine-0.1.0.tar.gz
- Upload date:
- Size: 14.6 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.14.6
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
a7ba47001a6d8cdf5b5655c836ed45e267fab4e535433c0165f735d05898fcd1
|
|
| MD5 |
79374d7b23e6ce39e66c1bc1a7fcffda
|
|
| BLAKE2b-256 |
f9b0e180df8ddfa19b59f8612ebe917ab946b9b9e6755e998c8c494b73057539
|
File details
Details for the file lac_engine-0.1.0-py3-none-any.whl.
File metadata
- Download URL: lac_engine-0.1.0-py3-none-any.whl
- Upload date:
- Size: 10.0 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.14.6
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
6a4150673ec049a10e3b31ff4dc706298e991c596a08d29e20ce5f943f995760
|
|
| MD5 |
050e6796fd4497fcbda2721dddcc40c9
|
|
| BLAKE2b-256 |
548c2a8fc3082149c231a668a68ce114a0bf3c535e7bd75c0cbad7cfc1533819
|