Draft v1.0 Open Standard

Agentic
Collaboration
Standard

One folder. Any agent. Any tool.

ACS defines a unified .agents/ folder for agent-ready projects. Write your context, skills, and permissions once — use them across Claude Code, Cursor, Copilot, and any future tool.

Get started → Read the spec
.agents/acs.yaml
# ACS Manifest v1.0
 
acs_version: "1.0"
 
project:
name: "my-app"
description: "E-commerce platform"
framework: nextjs
 
layers:
context: true
skills: true
commands: true
agents: true
permissions: true
 
compatible_with:
- claude-code
- cursor
- any
context ·skills ·commands ·agents ·permissions ·progressive disclosure ·vendor neutral ·open standard ·SKILL.md compatible ·MCP ready · context ·skills ·commands ·agents ·permissions ·progressive disclosure ·vendor neutral ·open standard ·SKILL.md compatible ·MCP ready ·
Works with
Claude CodeCursorGitHub CopilotAiderOpenAI CodexGemini CLIany agent →
The problem

Every tool reinvents
the same wheel.

Agent configuration is fragmented. Your context, skills, and rules live in vendor-specific formats that don't travel.

Without ACS
⚠  CLAUDE.md for Claude Code only
⚠  .cursorrules for Cursor only
⚠  No shared permissions model
⚠  Skills defined per-tool, not portable
⚠  Switch tools → lose everything
⚠  No standard for non-dev users
With ACS
 One .agents/ folder for everything
 Same config across all tools
 Explicit, version-controlled permissions
 SKILL.md-compatible reusable skills
 Portable when you switch tools
 Human-readable for non-devs too
Specification

Five layers.
One folder.

ACS is additive. Start with context only. Add layers as your project grows.

📄

context/

What agents need to know. Architecture, conventions, do-not-touch rules.

v1.0

skills/

Reusable capabilities. SKILL.md compatible with agentskills.io.

v1.0

commands/

Named, single-shot tasks. Like slash commands for any agent.

v1.0
🤖

agents/

Named subagents: reviewer, tester, docs-writer, and more.

v1.0
🔒

permissions/

Explicit allow/deny rules for file access. Committed to git.

v1.0

workflows, hooks, profiles

Multi-step orchestration and lifecycle triggers.

v2.0
📁 your-project/
├─ AGENTS.md # human overview
└─ .agents/    # ACS root
├─ acs.yaml    # manifest
├─ context/
└─ project.md
├─ skills/
└─ create-component/
├─ SKILL.md
└─ references/
├─ commands/
└─ explain-error.md
├─ agents/
└─ reviewer.md
└─ permissions/
└─ policy.yaml
How it works

Progressive disclosure.
Load only what you need.

ACS uses a three-tier loading strategy. Agents get the right context at the right time — without bloating the session.

Tier 01

Always in context

Manifest, permissions, and project context load at session start. Skill names only — ~50 tokens each.

~200 tokens total overhead
Tier 02

On activation

Full SKILL.md body loads when the agent decides the task matches. Commands and agents load on invocation.

< 5,000 tokens per skill
Tier 03

On demand

Reference files, scripts, and assets load only when the skill instructions reference them explicitly.

zero cost if unused
Compatibility

Designed to coexist,
not replace.

ACS works alongside every existing standard. Add .agents/ to any project without breaking what's there.

AGT

AGENTS.md

Keep your AGENTS.md for broad compatibility. ACS adds structured layers on top. They live side by side in the same repo.

SKL

SKILL.md (agentskills.io)

ACS skills use the exact same SKILL.md format. Any agentskills.io skill works inside .agents/skills/ with zero changes.

MCP

Model Context Protocol

MCP handles runtime tools. ACS handles project configuration. Different layers, fully complementary.

CLD

CLAUDE.md

Migrate incrementally. The planned acs compile CLI will generate CLAUDE.md from ACS automatically.