English | 中文 | 日本語 | 한국어

🤝 collab-cli

A collaboration protocol that makes multiple AI agents work together like a real team

Works with Claude Code, Reasonix, Codex, WorkBuddy, Cursor, or any AI agent — on one device or across a LAN.

---

Table of Contents

• What is this?
• Two modes: Single device vs Multi-device
• How it works (protocols & principles)
• Quick start
• Core concepts
• Complete walkthrough
• Agent integration
• LAN Node (cross-device)
• CLI reference
• Architecture
• Development

---

What is this? {#what-is-this}

collab-cli solves a simple problem: multiple AI agents working on the same project don't know what each other are doing.

Without collab-cli:

Claude Code session 1  ──→  Changes the API
Claude Code session 2  ──→  Doesn't know, changes it again
WorkBuddy              ──→  Runs nightly job, doesn't see the change
You (the human)        ──→  Explains everything from scratch every time

With collab-cli:

Claude Code  ──→  Writes to .shared/SHARD.md: "API changed to v2"
WorkBuddy    ──→  Reads SHARD.md: "Oh, API changed, I'll update my job"
You          ──→  Don't need to explain anything

It's a shared folder (.shared/) + a protocol that all agents follow. That's it. No servers required for single-device use.

Two modes: Single device vs Multi-device {#two-modes}

collab-cli works in two modes. Choose based on where your agents run:

Mode 1: Single device (most users)

Your computer
├── .shared/              ← One shared folder, all agents read/write here
│   ├── SHARD.md             Current project state (≤80 lines)
│   ├── BADGE-claude-01.md   Claude's identity & permissions
│   ├── BADGE-workbuddy-01   WorkBuddy's identity & permissions
│   ├── inbox/               Messages between agents
│   ├── tasks/               Task board
│   └── memory/              Knowledge fragments
├── Claude Code           ← Reads .shared/
├── WorkBuddy             ← Reads .shared/
└── Codex                 ← Reads .shared/

When to use: All your agents run on the same computer. This is the most common case.

How it works: Pure filesystem. Agents read/write files in .shared/. Zero network, zero servers, zero configuration.

Mode 2: Multi-device (LAN)

Computer A (192.168.1.100)           Computer B (192.168.1.101)
├── .shared/                         ├── .shared/
│   ├── SHARD.md  ←──── sync ────→  │   ├── SHARD.md
│   ├── tasks/   ←──── sync ────→   │   ├── tasks/
│   ├── inbox/codex-1/ (own only)   │   ├── inbox/codex-2/ (own only)
│   └── BADGE-*.md                  │   └── BADGE-*.md
├── collab node :9527  ←── HTTP ──→ ├── collab node :9527
└── Codex-1                          └── Codex-2

When to use: Your agents run on different computers on the same WiFi/LAN.

How it works: Each computer runs a collab node (lightweight HTTP server). Nodes discover each other via UDP broadcast. SHARD.md and tasks sync every 10 seconds. Inbox messages push in real-time.

Which mode should I use?

Your setup	Mode	Command
All agents on one computer	Single device	collab setup --devices 1
Agents on 2+ computers, same LAN	Multi-device	collab setup --devices 2
Agents on different networks	Not supported yet	Use git sync as workaround

How it works (protocols & principles) {#how-it-works}

Communication protocols

collab-cli uses four different protocols depending on what's happening:

Protocol	Port	Used for	When
Filesystem	N/A	SHARD, tasks, badges, memory	Single-device mode (always)
UDP Broadcast	9528	Peer discovery	Multi-device mode only
HTTP REST	9527	Message routing, SHARD sync, tasks sync	Multi-device mode only
MCP (JSON-RPC)	stdio	Plugin integration with Reasonix/Claude Desktop	Optional

Core principle: Files are the protocol

The entire system is built on plain Markdown files with YAML frontmatter. Any agent that can read and write files can participate. No special SDK, no API client, no runtime dependency.

---
id: MSG-001
from: claude-01
to: workbuddy-01
priority: P1
status: unread
---

# Please review the login module

The code is at src/auth/login.py.

This is what a message looks like. It's just a file. Any text editor can read it. Any AI agent can parse it.

Core principle: Role-based access control

Every agent gets a badge when joining a project. The badge defines what the agent can and cannot do:

L4 Chief Engineer ──┬── Full access (read/write everything)
                    │
L3 Reviewer ────────┤── Can approve tasks, write SHARD
                    │
L2 Contributor ─────┤── Can write memory, submit reviews
                    │
L1 Executor ────────┤── Can write own tasks + inbox
                    │
L0 Observer ────────┴── Read-only

Different sessions of the same agent can hold different badges. Badges expire when the session ends.

Core principle: 3-layer memory with auto-decay

To prevent memory bloat (every agent reading hundreds of lines), memory is split into three layers:

L0  SHARD.md     ← "What's true right now" (≤80 lines, every agent reads this)
L1  memory/       ← "What we've decided" (by topic, ≤50 lines each)
L2  archive/      ← "What happened before" (by date, only consulted when needed)

When SHARD exceeds 80 lines, old entries automatically move to archive/. New agents only need to read 80 lines to understand the full picture.

Core principle: Task state machine

Every task follows a strict lifecycle:

DRAFT → ASSIGNED → IN_PROGRESS → REVIEW → DONE
                                  ↓
                               REWORK → IN_PROGRESS

Tasks cannot skip REVIEW. The Chief Engineer reviews, then the human gives final approval.

Core principle: Handshake on entry

Every agent must execute a handshake when entering a project:

Step 1: Read MANIFEST.md   → System rules
Step 2: Read SHARD.md      → Current state (≤80 lines)
Step 3: Read BADGE-{id}.md → Your identity & permissions
Step 4: Check inbox        → Unread messages
Step 5: Check tasks        → Your active tasks
Step 6: Output summary     → Then respond to user

This ensures no agent starts working without knowing the current context.

Quick start {#quick-start}

Option A: Setup wizard (recommended)

npm i -g collab-cli

# Single device — just works
collab setup --devices 1 --project "My Project"

# Multi device — generates per-device instructions
collab setup --devices 2 \
  --project "My Project" \
  --device-1 "DeviceA:codex-1@Codex" \
  --device-2 "DeviceB:codex-2@Codex"

The wizard will:

1. Initialize .shared/ directory with all required files
2. Issue badges for all agents
3. Generate per-device startup instructions
4. Create peers.yaml with LAN config (multi-device only)

Option B: Manual setup

npm i -g collab-cli

# Initialize
cd my-project
collab init --project "My Project"

# Issue badges
collab badge issue claude-01 --role L4 --assigned-by user
collab badge issue workbuddy-01 --role L2 --assigned-by user

# Set up agent instructions
cat node_modules/collab-cli/src/templates/CLAUDE_PROTOCOL.md >> .claude/CLAUDE.md

# Start collaborating
collab task create "First task" --assignee claude-01 --priority P0
collab inbox send --from claude-01 --to workbuddy-01 --title "Hello" --priority P1

Core concepts {#core-concepts}

🪪 Badge — Agent identity & permissions

Each agent receives a badge when entering a project. The badge defines the agent's role, permissions, and scope.

Level	Name	Can do	Cannot do
L0	Observer	Read SHARD, memory	Write anything
L1	Executor	Write own tasks + inbox	Write SHARD, approve tasks
L2	Contributor	L1 + write memory + submit reviews	Approve tasks, manage badges
L3	Reviewer	L2 + approve tasks + write SHARD	Change other badges
L4	Chief Engineer	Everything	—

📬 Inbox — Structured messaging

Agents communicate through structured messages with priority (P0-P3), type (task/review/approval/question), and optional response-required flag.

📋 Task — Lifecycle-managed work items

Tasks follow a state machine: DRAFT → ASSIGNED → IN_PROGRESS → REVIEW → DONE. Only the assignee can change status. Completion requires Chief Engineer review + user confirmation.

🧠 Memory — Three layers with auto-decay

L0 SHARD (live, ≤80 lines) → L1 memory (fragments, ≤50 lines each) → L2 archive (compressed by date). Prevents memory bloat while preserving history.

🤝 Handshake — Automatic context loading

Every agent reads SHARD + badge + inbox + tasks on startup. No manual explanation needed.

Complete walkthrough {#complete-walkthrough}

See the real-world walkthrough in the detailed READMEs:

• English walkthrough
• 中文演练

Agent integration {#agent-integration}

Agent	Integration method	File
Claude Code	Append to .claude/CLAUDE.md	CLAUDE_PROTOCOL.md
Reasonix	Copy to .reasonix/system.md or MCP plugin	REASONIX_PROTOCOL.md
WorkBuddy	Append to .workbuddy/memory/MEMORY.md	AGENT_PROTOCOL.md
Cursor	Merge into .cursor/rules	CURSOR_PROTOCOL.md
Codex	Copy as AGENTS.md	CODEX_PROTOCOL.md
Any agent	Place AGENT_PROTOCOL.md in project root	AGENT_PROTOCOL.md

See templates/ for all protocol files.

LAN Node (cross-device) {#lan-node}

How multi-device sync works

┌──────────────────────────────────────────────────────────────────┐
│                      Multi-Device Architecture                    │
│                                                                  │
│  Device A                                    Device B            │
│  ┌─────────────────┐                       ┌─────────────────┐  │
│  │ collab node     │   UDP Broadcast       │ collab node     │  │
│  │ HTTP :9527      │◄─────────────────────►│ HTTP :9527      │  │
│  │ UDP  :9528      │   (auto-discovery)    │ UDP  :9528      │  │
│  └────────┬────────┘                       └────────┬────────┘  │
│           │                                         │           │
│           │            SHARD sync (10s)              │           │
│           │◄──────────── HTTP push ─────────────────►│           │
│           │            Tasks sync (10s)              │           │
│           │◄──────────── HTTP push ─────────────────►│           │
│           │            Inbox (real-time)             │           │
│           │◄──────────── HTTP push ─────────────────►│           │
│           │                                         │           │
│  ┌────────┴────────┐                       ┌────────┴────────┐  │
│  │ .shared/        │                       │ .shared/        │  │
│  │ SHARD.md ✅ sync│                       │ SHARD.md ✅ sync│  │
│  │ tasks/ ✅ sync  │                       │ tasks/ ✅ sync  │  │
│  │ inbox/a1 ❌ own │                       │ inbox/a2 ❌ own │  │
│  │ memory/ ✅ sync │                       │ memory/ ✅ sync │  │
│  └─────────────────┘                       └─────────────────┘  │
│                                                                  │
│  Codex-1                                   Codex-2              │
└──────────────────────────────────────────────────────────────────┘

What syncs and what doesn't

File	Synced?	Protocol	Strategy
SHARD.md	✅	HTTP push every 10s	Version-based: newer version wins
tasks/	✅	HTTP push every 10s	Status merge: more advanced status wins
memory/	✅	HTTP push	Full replace
inbox/	❌	—	Per-device by design (each device stores its own agent's messages)
MANIFEST.md	—	Manual	Same on all devices (keep in sync via git)
BADGE-*.md	—	Manual	Same on all devices

LAN discovery protocol

1. Each node broadcasts a UDP packet on port 9528 every 5 seconds
2. The packet contains: { nodeId, agents: [...], apiPort }
3. Other nodes receive the broadcast and register the peer
4. After 15 seconds without a heartbeat, the peer is marked as offline

First-time setup on a new device

When a new device joins the network, it needs to pull existing data:

# On the new device
collab node pull --host 192.168.1.100 --port 9527 --token <token>

This pulls the current SHARD.md and all tasks from the existing node.

CLI reference {#cli-reference}

# ── Setup (start here) ──
collab setup                                    Guided setup wizard
collab setup --devices 1 --project "Name"       Single device
collab setup --devices 2 --device-1 "A:agent@Type" --device-2 "B:agent@Type"

# ── System ──
collab init --project "Name"                    Initialize .shared/
collab status                                   Global status overview
collab handshake <agent-id>                     Agent handshake check

# ── Badges ──
collab badge issue <id> --role <L0-L4>          Issue badge
collab badge show <id>                          Show badge details
collab badge list                               List all badges

# ── Tasks ──
collab task create <title> --assignee <id>      Create task
collab task list [--status <s>] [--assignee <id>]
collab task status <id>                         Task details
collab task update <id> <status>                Update status

# ── Inbox ──
collab inbox check <id>                         Check unread messages
collab inbox send --from <id> --to <id>         Send message
collab inbox read <id> <msg-id>                 Read (mark as read)
collab inbox done <id> <msg-id>                 Mark as done

# ── Memory ──
collab memory compact                           Auto-archive old entries
collab memory stats                             Layer statistics

# ── Heartbeat ──
collab heartbeat <id>                           Start persistent monitoring
collab heartbeat <id> --once                    Single check (exit 2 = P0/P1)

# ── LAN Node ──
collab node start [--agents <ids>] [--port N]   Start LAN node
collab node pull --host <ip>                    Pull SHARD + tasks from peer
collab node status                              Show node + discovered peers

# ── MCP Server ──
collab mcp                                      Start MCP server (stdio)

Architecture {#architecture}

File structure

.shared/
├── MANIFEST.md              System declaration + agent registry + role definitions
├── SHARD.md                 L0 live memory (≤80 lines) — current state
├── BADGE-{agent-id}.md      Per-agent badge (multi-badge, one per agent)
├── peers.yaml               LAN node configuration (multi-device only)
├── inbox/{agent-id}/        Message inbox (per-agent directory)
│   └── 001-{topic}.md       Structured message (YAML frontmatter + Markdown body)
├── tasks/T-xxx.md           Task file (state machine + progress log)
├── memory/                  L1 memory fragments (by topic, ≤50 lines each)
│   ├── decisions.md
│   ├── lessons.md
│   └── architecture.md
├── archive/                 L2 archive (by date, auto-compressed)
│   └── 2026-06-06.md
└── conflicts/               Conflict records (awaiting Chief Engineer arbitration)
    └── C-{timestamp}.md

Module structure

collab-cli/
├── bin/collab.js              CLI entry (20+ subcommands)
├── src/
│   ├── commands/              Command implementations
│   │   ├── init.js            Initialization + migration
│   │   ├── setup.js           Setup wizard
│   │   ├── status.js          Global status
│   │   ├── badge.js           Badge management
│   │   ├── task.js            Task lifecycle
│   │   ├── inbox.js           Messaging
│   │   ├── memory.js          Memory decay
│   │   ├── conflict.js        Conflict resolution
│   │   ├── heartbeat.js       Heartbeat monitoring
│   │   ├── node.js            LAN node command
│   │   └── mcp-server.js      MCP server
│   ├── core/                  Core modules
│   │   ├── protocol.js        Handshake protocol + permissions
│   │   ├── shard.js           SHARD management + auto-archive
│   │   ├── fs-lock.js         Optimistic file locking
│   │   └── yaml.js            YAML frontmatter engine
│   ├── node/                  LAN node modules
│   │   ├── discovery.js       UDP broadcast peer discovery
│   │   ├── server.js          HTTP API server
│   │   ├── router.js          Message routing (local/remote)
│   │   └── sync.js            Cross-device SHARD/tasks sync
│   ├── templates/             Protocol templates for each agent type
│   └── utils/                 Timestamp + markdown utilities
├── package.json               collab-cli
├── README.md                  English
├── README.zh-CN.md            Chinese
├── README.ja.md               Japanese
└── README.ko.md               Korean

Conflict resolution (three layers)

Layer 1: Prevention
├─ Task assignee is explicit → others don't touch it
├─ SHARD.md checks last_updated_by before writing
└─ Scope limits each agent's working directories

Layer 2: Detection
├─ Frontmatter last_updated_by + last_updated_at
└─ Optimistic lock compares version on write

Layer 3: Arbitration
├─ Conflict written to conflicts/C-{timestamp}.md
├─ Chief Engineer resolves within 24h
└─ If Chief Engineer can't decide → escalate to user

Comparison with alternatives

Feature	Manual coordination	Git branches	collab-cli
Real-time messaging	❌	❌	✅
Role-based permissions	❌	Branch-level	File-level
Shared memory	Verbal	Commit messages	Structured 3-layer
Task lifecycle	Verbal	PR/Issue	Built-in state machine
New agent onboarding	Re-explain everything	Clone repo	Handshake auto-aligns
Cross-device sync	N/A	git pull/push	Auto HTTP sync (10s)
Token cost	Re-read everything	N/A	≤80 lines live memory

Development {#development}

git clone https://github.com/yinsang0910-star/collab-cli.git
cd collab-cli
npm install
npm test          # 95 tests passing
npm link          # Local development

License

MIT

---

If this project helped you, give it a ⭐!