A file-based + TCP real-time protocol enabling two AI coding assistants to collaborate on the same project. Claude Code and OpenAI Codex (or VS Code Copilot) share messages, tasks, handoffs, and knowledge — no server, no network, just a .a2cp/ directory.
No pip install and no build step - Python 3.8+ and the standard library are all it needs.
git clone https://github.com/TOSKA1226/a2cp-claudecode-codex-.git
cd a2cp-claudecode-codex-
python a2cp.py --help # smoke test# 1. Enter your project
cd /path/to/your-project
# 2. Initialize A2CP
python /path/to/a2cp-claudecode-codex-/a2cp.py init
# 3. Launch the interactive Command Center (browser)
python /path/to/a2cp-claudecode-codex-/a2cp.py center
# 4. Or use CLI commands
python a2cp.py bridge start # Start real-time bridge
python a2cp.py task create "My task" # Create a task
python a2cp.py send --to codex # Send a message
python a2cp.py dashboard # Terminal live panelA2CP lets two AI coding assistants collaborate like a human dev team:
| Feature | How It Works |
|---|---|
| Messages | JSON files in .a2cp/mailbox/{ai}/inbox/ + optional TCP real-time delivery |
| Task Board | Kanban: backlog → todo → in-progress → review → done |
| Handoffs | Structured JSON with decisions, blockers, and context for the receiver |
| Shared Memory | 7-category knowledge base (architecture, conventions, decisions, patterns, gotchas, API reference, dependencies) |
| Human Notes | .a2cp/NOTES.md — human-editable markdown both AIs read |
| Live Dashboard | Terminal TUI with keyboard shortcuts |
| Web Command Center | Browser-based control panel with one-click task assignment, message sending, and memory management |
.a2cp/
├── config.json # Protocol configuration (v2.0)
├── NOTES.md # Human-editable shared notes
├── mailbox/
│ ├── claude/ (inbox/ read/ sent/)
│ └── codex/ (inbox/ read/ sent/)
├── tasks/ (backlog/ todo/ in-progress/ review/ done/)
├── handoffs/ # Work handoff records
├── memory/ # Shared knowledge (category JSON files)
├── locks/ # File-based mutexes
└── cli/ # Local launcher scripts
# Core
a2cp init # Initialize .a2cp/ in current project
a2cp status # Text overview
a2cp dashboard # Live terminal TUI
a2cp center # Web Command Center (browser)
# Bridge (real-time)
a2cp bridge start # Start TCP bridge server
a2cp bridge stop # Stop bridge server
a2cp bridge status # Check bridge status
a2cp bridge ping # Heartbeat — check if other AI is online
a2cp bridge monitor # Live message feed
# Messages
a2cp check --ai claude # Check inbox
a2cp read --all --ai claude # Read all messages
a2cp send --to codex --type task_request --subject "Title" --body "..."
a2cp reply <msg_id> --body "..."
# Tasks
a2cp task create "Title" --desc "..." --priority high
a2cp task list --status backlog
a2cp task claim T-0001 --ai claude
a2cp task start T-0001
a2cp task review T-0001
a2cp task approve T-0001
# Handoffs
a2cp handoff create --to codex --done "What was done" --next "What's next"
a2cp handoff list
a2cp handoff ack H-0001
# Memory
a2cp memory set <key> <value> --category architecture
a2cp memory get <key>
a2cp memory search "query"
a2cp memory list --category architecture
# Notes
a2cp notes read
a2cp notes add --text "..."
a2cp notes edit- Python 3.8+ — zero dependencies beyond stdlib
- Windows / macOS / Linux — cross-platform
- Modern terminal — for ANSI dashboard (Windows Terminal, iTerm2, etc.)
- Human initializes the project and starts the bridge
- Claude creates tasks, writes shared memory, and reviews code
- Codex checks inbox, claims tasks, writes code, submits for review
- Both AIs communicate via messages and handoffs
- Human orchestrates everything from the Command Center
Found a bug or have an idea? Open an issue or PR. See PROTOCOL.md for the full protocol spec.
MIT — see LICENSE