0xNyk/lacp

What it solves

LACP provides a local control plane and execution wrapper for coding agents (such as Claude, Codex, and Hermes). It solves the problem of running autonomous agents with insufficient oversight by introducing deterministic routing, approval gates, and evidence-based verification, ensuring that agent actions are governed by local policy rather than being fully autonomous and unmonitored.

How it works

LACP acts as a layer between the agent host and the machine. It intercepts agent invocations and routes them through a series of checks: risk tier selection (safe, review, or critical), budget ceilings, and context contract validation. It uses a hook pipeline (specifically for Claude Code) to block dangerous operations and validate output quality. To maintain organization and safety, it isolates parallel agent sessions using Git worktrees and tmux/dmux sessions. It also implements a multi-layer memory stack—including an Obsidian-based knowledge graph—to maintain persistent agent identity and provenance.

Who it’s for

Developers who require reproducible local agent runs with explicit pass/fail gates and auditable artifacts, and who prefer local policy control over managed cloud services.

Highlights

• Policy-Gated Execution: Routes commands through risk tiers and budget ceilings to prevent unauthorized or expensive operations.
• Evidence Pipelines: Generates machine-verifiable evidence (e2e, API, and contract tests) to validate changes before they are promoted to a PR.
• 5-Layer Memory Stack: Combines session memory, an Obsidian knowledge graph, ingestion pipelines, and code intelligence for persistent context.
• Runtime Isolation: Uses Git worktrees and session fingerprints to isolate parallel agents and prevent state corruption.
• Claude Code Hooks: Modular Python hooks that enforce quality gates and block dangerous system calls like chmod 777 or fork bombs.