AOG — Multi-Agent CLI Orchestrator

AOG (Anthropic, OpenAI, Google) is an open-source MCP server that orchestrates Claude Code, Codex CLI, and Gemini CLI as a collaborative multi-agent coding team. Multiple models work the same problem independently, then cross-review and synthesize — applied to CLI coding agents working on real code.

block-beta
  columns 3

  space Client["MCP Client"] space
  space:3
  space AOG["AOG Server — Router • Orchestrator • State"] space
  space:3
  Claude["Claude Code\nPlanning"] Codex["Codex CLI\nSpeed"] Gemini["Gemini CLI\nResearch"]
  space:3
  space Review["Cross-Review • Scoring • Synthesis"] space

  Client --> AOG
  AOG --> Claude
  AOG --> Codex
  AOG --> Gemini
  Claude --> Review
  Codex --> Review
  Gemini --> Review

  style Client fill:#1e293b,stroke:#334155,color:#e2e8f0
  style AOG fill:#2d3748,stroke:#4a5568,color:#e2e8f0
  style Claude fill:#d97706,stroke:#b45309,color:#fff
  style Codex fill:#059669,stroke:#047857,color:#fff
  style Gemini fill:#2563eb,stroke:#1d4ed8,color:#fff
  style Review fill:#7c3aed,stroke:#6d28d9,color:#fff

Loading

See It In Action

A single council_run dispatched the same task to all three agents in parallel worktrees:

Agent	What It Built
Claude	Clean 9-line implementation using Date.now()
Codex	13-line version with a HealthStatus TypeScript interface and process.uptime()
Gemini	12-line version with JSDoc documentation and process.uptime()

All three implementations were anonymized, cross-reviewed, and scored. The chairman merged the best parts — Codex's type interface with Gemini's docs. Total time: ~82 seconds.

Prerequisites

• Node.js >= 20
• Git (for worktree isolation)
• At least one of:
  • Claude Code (Claude Max subscription)
  • Codex CLI (ChatGPT Plus/Pro subscription)
  • Gemini CLI (Google account, free tier available)

Don't have all three? AOG works with any subset — even just one CLI. Delegate mode routes tasks to whatever you have installed. Multi-agent modes unlock when you have two or more.

Setup

1. Verify CLIs are installed

claude --version   # Need 2.1+
codex --version    # Need 0.106+
gemini --version   # Need 0.30+

2. Verify each CLI is authenticated

Run these one at a time. Each should return JSON and exit cleanly.

claude -p "say hi" --output-format json --max-turns 1
codex exec "say hi" --full-auto --json
gemini -p "say hi" --output-format json --yolo

If Gemini asks for auth, run gemini interactively first to complete Google OAuth.

3. Install and register AOG

Your project must be a git repository. AOG uses git worktrees for agent isolation. Run git init if needed.

Option A — Install from npm:

claude mcp add --transport stdio aog -- npx -y aog-mcp-server

Option B — Run from source:

git clone https://github.com/LonglifeIO/AOG.git
cd AOG
npm install
npm run build
claude mcp add --transport stdio aog -- node /path/to/AOG/dist/index.js

4. Verify AOG is registered

claude mcp list   # Should show 'aog'

5. First test

Ask Claude Code:

"Use AOG to have codex say hello world"

If that works, try a real task:

"Use AOG to synthesize research and build a rate limiter for the API endpoints"

Three Modes

DELEGATE — Single Agent Routing

Routes each task to the best available CLI:

Task Type	Preferred Agent	Why
IMPLEMENT	Claude Code	Best multi-file planning
RESEARCH	Gemini CLI	1M token context window
GENERATE	Codex CLI	Fastest execution
REVIEW	Gemini CLI	Large context for diffs
DEBUG	Claude Code	Diagnostic reasoning

COUNCIL — Multi-Agent Consensus

1. Each agent gets an isolated git worktree
2. All agents implement the task simultaneously
3. Tests run in each worktree
4. Agents cross-review anonymized diffs (Implementation A/B/C)
5. Chairman merges the best parts into the final implementation

PIPELINE — Staged Workflows

Chain specialized agents across multiple stages:

Template	Stages
full-council	implement → test → review → synthesize → final test
quick-fix	fix → test → review
migration	research → plan → approve → implement → test → review → synthesize
dependency-update	analyze → update → test → review
research-synthesis	synthesize research → plan → approve → build

The research-synthesis pipeline is the simplest way to start: drop research outputs from multiple LLMs into a research/ folder, then let AOG synthesize and build from the plan.

Token Efficiency

AOG is designed for token-constrained environments (Claude Pro, API budgets).

• task_file param — All tools accept task_file as an alternative to task. Write your spec to a temp file and pass the path instead of inlining thousands of characters. Same result, saves the caller's output tokens.
• Compact responses — Tool responses contain only taskId, status, summary, files_changed, duration_ms, and session_path. Full diffs, reviews, and synthesis details go to .aog/sessions/ on disk. Use council_status with detail_level: "diffs" or "full" to retrieve more.
• Lean prompts — Worker instructions, cross-review prompts, and synthesis prompts are minimized. Reviews use git diff --stat instead of full diffs.
• Debug logging — Set AOG_DEBUG_TOKENS=true to log token estimates (chars/4) per session in .aog/sessions/{taskId}.json.

Progress Notifications

Long-running operations push real-time MCP progress notifications so you can see what's happening instead of staring at a spinner:

Council started — 2 agents (claude, gemini)
Worktrees created for claude, gemini
claude: implementing…
gemini: implementing…
gemini: done (72s, 4 files changed)
claude: done (108s, 3 files changed)
Running tests…
Tests: 2/2 agents passed
Cross-review started (claude, gemini)
Cross-review complete
Chairman claude synthesizing…
Synthesis complete (best-wins)

Progress is reported for all three modes (delegate, council, pipeline) at each phase: worktree creation, agent spawn/complete, tests, cross-review, and synthesis.

Submodule Support

Git worktrees automatically run git submodule update --init --recursive after creation, so projects with submodules (e.g. shared libraries tracked as separate repos) work out of the box.

Security

• Spawns official CLI binaries only — no OAuth token extraction, no SDK auth hacks
• Git worktree isolation — each agent works in its own directory
• Output sanitization — strips prompt injection patterns between agents
• Approval gates — configurable human-in-the-loop at critical pipeline stages
• Audit logging — all operations logged to .aog/sessions/

Configuration

Create aog.config.yaml in your project root:

agents:
  claude:
    enabled: true
    model: sonnet
    maxTurns: 20
    maxBudgetUsd: 5.0
  codex:
    enabled: true
    model: gpt-5.4
  gemini:
    enabled: true
    model: gemini-2.5-pro

defaults:
  chairman: claude
  timeout: 300000
  pipeline: full-council

security:
  require_approval_before_merge: true
  allow_permission_bypass: true
  max_parallel_agents: 3
  max_budget_per_task_usd: 10.0
  audit_logging: true

Development

git clone https://github.com/LonglifeIO/AOG.git
cd AOG
npm install
npm run build   # Compile TypeScript
npm run dev     # Start MCP server with tsx

Not Yet Built

• aog init setup wizard
• Container isolation mode
• Pipeline resume after approval pause
• Metrics / learned adaptive routing

Contributing

This project is in early development. Issues, bug reports, and PRs welcome. See docs/decisions/ for architecture context before making major changes.

Troubleshooting

• "MCP failed to reconnect" → Rebuild with npm run build, then re-register with claude mcp add
• Gemini auth error → Run gemini interactively to complete Google OAuth, then retry
• Codex timeout → Check your ChatGPT Plus/Pro subscription is active
• AOG works with any subset of CLIs — you don't need all three installed

License

MIT