A new open-source project called Mati has emerged on Hacker News, offering a local guardrail layer designed to give developers more granular control over AI coding agents. The tool specifically targets Claude Code and Codex, addressing a critical pain point in agentic workflows: the gap between what an LLM can technically do and what a product team actually allows.

The Tacit Knowledge Problem

The creator of Mati, identified on GitHub as ioni-dev, highlights a persistent issue in software engineering: many important product constraints exist only in the heads of engineers and product managers. These tacit rulesβ€”ranging from architectural boundaries to specific security protocolsβ€”are often lost in translation when handed off to autonomous coding agents. Without explicit enforcement, these agents can inadvertently violate core product principles, leading to technical debt or broken features.

Local Enforcement, Not Just Prompts

While prompt engineering is the standard approach to constraining LLMs, Mati takes a different tack by acting as a local enforcement layer. This suggests a shift from hoping the model understands the instructions to actively filtering or blocking actions that violate defined rules. By operating locally, Mati likely intercepts the agent's proposed actions or code changes before they are fully executed or committed, providing a safety net that is independent of the model's probabilistic nature.

Key Takeaways

  • Mati is an open-source guardrail layer for Claude Code and Codex.
  • It aims to enforce tacit product constraints that usually live in engineers' heads.
  • The tool provides a local control mechanism, moving beyond simple prompt-based restrictions.
  • The project was submitted to Hacker News on September 19, 2026.

The Bottom Line

As coding agents become more autonomous, the industry needs robust guardrails that go beyond prompt engineering. Mati's approach of local, deterministic enforcement is a necessary evolution for enterprise-grade AI coding tools.

Technical Context

The source material is limited, as the Hacker News post is brief and the GitHub repository content was truncated during retrieval. However, the core proposition is clear: Mati serves as a middleware or wrapper that applies specific rules to the output or actions of Claude Code and Codex. This aligns with the broader industry trend of building 'sandboxed' or 'governed' AI agents, where the focus shifts from raw capability to controlled, reliable autonomy. The lack of detailed technical documentation in the initial post suggests the project is likely in an early stage, possibly a proof-of-concept or a niche utility for power users who are frustrated by the 'hallucination' of constraints by large models.