Skip to content

Welcome to Lár: The Glass Box Agent Framework

Lár (Irish for "core" or "center") by SnathAI™ is the open-source "PyTorch for Agents."

It is a "define-by-run" framework for building auditable, reliable, and production-ready AI agents.

Lár is engineered as a direct solution to the "black box" problem. While other frameworks hide their logic in complex, "magic" executors that are impossible to debug, lar is a simple, "dumb" engine. It runs one node at a time, logs exactly what happened, and then moves on.

This "glass box" philosophy gives you a complete, step-by-step "flight data recorder" for every agent run, allowing you to build systems you can actually trust.

Why We Built Lár?

The Problem: Compliance is Impossible with Black Boxes

You are a developer building an agent for Clinical Trials or Government Intelligence. - The FDA requires 21 CFR Part 11 audit trails. - The SCIF has no internet.

Traditional frameworks fail both tests. They are non-deterministic "chatbots" that require cloud connectivity for tracing.

The Lár Solution: The First GxP-Ready Framework

We built the Lár Engine to bring the Scientific Method to AI Agents.

Lár is: 1. Deterministic: Same seed + Same Graph = Identical Execution. 2. Auditable: Generating a forensic flight log is the default behavior. 3. Air-Gap Capable: Build on a laptop -> Serialize to JSON -> Run offline.

The Lár Solution: Your Agent's Flight Recorder

We built the Lár Engine as a direct antidote to the "Black Box Tax."

Lár's core output is not just an answer; it is a full, immutable Flight Log.

Lár's "define-by-run" architecture forces transparency: the GraphExecutor is designed as a simple generator that runs one node, logs the exact state change, and then pauses, producing a verifiable record for every step.

This means you can always:

  • Pinpoint Failure: See the exact node (ToolNode) that failed, the exact error (e.g., APIConnectionError), and the full state of the agent at the moment of collapse.

  • Audit Costs: Track token usage per node, ensuring you can justify and optimize every single API call.

  • Build Trust: Move your agent from a chaotic chat loop to a predictable, testable assembly line that your team and your company can actually trust in production.

We are not selling magic; we are selling trust.

Stop guessing. Start building agents you can trust!

Get Started in 3 Minutes https://docs.snath.ai/getting-started/