"Don't generate code. Orchestrate it."
ai-context-os is an installable LLM Orchestrator Framework designed to manage the intelligence, constraints, and context of AI Agents (Cursor, Claude, Antigravity) within any repository. It treats "AI Context" as Infrastructure-as-Code.
- Recursive Evolution: The OS is now "living". Agents possess Legislative Rights (L4 Protocol) to autonomously detect knowledge gaps, research new technologies, and author new standardized skills into the OS without human intervention.
- Silent Bootstrapping: Zero-prompt initializations. Agents read the
.ultp_statebuffer to instantly sync environment states without running blocking shell commands. - Dynamic Shared Memory: Continuity between agent sessions through a centralized
.ai-context-os/memory/session.mdevent log. - Diamond Standards: Rigorous 90% test coverage threshold and 100% Purity checks for modular capabilities.
When working with Advanced Agents, context quickly degrades into a mess of conflicting rules, lost memories, and hallucinated patterns. ai-context-os solves this via a Fallback Architecture (Inheritance).
- Context Hygiene: Hides complex logic inside
.ai-context-os/and places tiny "pointer" files (.cursorrules,CLAUDE.md) in your root. - Enforces SSOT: Ensures all agents fall back to the unbreakable laws in
PROJECT_OS.mdif project-specific rules fail. - Self-Healing: Agents detect systemic failures and literally rewrite their own context rules to prevent a recurrence.
We organize intelligence into four distinct layers:
| Layer | Name | Description |
|---|---|---|
| L0 | Kernel | The immutable "Constitution" of your project (PROJECT_OS.md). |
| L1 | Adapters | Pointer files (.cursorrules, CLAUDE.md, GEMINI.md) that bridge the AI to the Kernel. |
| L2 | Skills | Modular capabilities (React, Fastify, TDD) automatically generated and vetted. |
| L3 | Memory | Operational session logs allowing continuity across distinct agent runs. |
├── PROJECT_OS.md # L0: The Kernel (Single Source of Truth)
├── CLAUDE.md # L1: Adapter for Claude/Antigravity
├── .cursorrules # L1: Adapter for Cursor AI
├── GEMINI.md # L1: Adapter for Gemini AI
├── skills/ # L2: Modular Capabilities (TDD, Frameworks)
└── memory/ # L3: Dynamic Shared Memory
The best way to leverage ai-context-os is to drop it into your existing projects to instantly structure their AI workflows.
Run the following in your project root to provision the hidden .ai-context-os/ folder and setup the L1 pointers:
npx ai-context-os .This keeps your root clean. Only .cursorrules, CLAUDE.md, and GEMINI.md are visible, acting as silent bootstrappers.
mkdir .ai-context-osin your project root.- Use this package as a template: Move
PROJECT_OS.mdand theskillsfolder into your.ai-context-os/directory. - Copy one of the
adapter-***.mdtemplates into your root (rename it to.cursorrulesorCLAUDE.md).
- Protocol-First: Rules in
PROJECT_OS.mdoverride any pre-trained AI assumptions. - Silent Synchronization: Agents establish context context entirely O(1) via the
.ultp_statecache without executing messy CLI commands. - Atomic Documentation: Every code change MUST include simultaneous documentation updates.
- Regression Assurance: Full test suites must be rerun after every modification.
| Command | Action |
|---|---|
npx ai-context-os . |
Quick Integration (Current Dir) |
npx ai-context-os audit --diamond |
Check Architectural Compliance against Purity rules |
npx ai-context-os scout |
Visualize the active Context Architecture and loaded skills |
This OS uses the Ultra-Low Token Protocol (ULTP) to serialize the entire system state into a tiny string:
[OS:A][L0:V;P:.ai-context-os/PROJECT_OS.md][L1:C,G,K][L2:tdd,fastify][M:V]
This reduces context overhead by 65%, providing high-density signaling for faster, cheaper, and more accurate AI orchestration.
The OS is designed to be self-writing. However, we welcome human PRs that improve the kernel or add new standardized L1 adapters. To trigger an AI-driven skill discovery in your fork, simply ask the agent to implement a technology it hasn't seen before.