Your GPUs are idle 70% of the time. CQ connects them to AI — zero config, any OS, encrypted.
| Without CQ | With CQ | |
|---|---|---|
| Your GPU at 2am | Idle, wasting electricity | Running your experiment |
| Cross-device AI | Loses context every session | Persistent memory everywhere |
| Code quality | Hope and pray | 6-axis automated review |
Start a training job from your laptop, phone, or any AI assistant. CQ's Hub dispatches tasks to whichever GPU is available, streams metrics back in real time, and stores artifacts automatically.
cq job submit --image pytorch --gpu 1 -- python train.pyEvery conversation contributes to your knowledge base. ChatGPT finds a bug root cause — Claude picks it up in the next session. Decisions, patterns, and discoveries persist across tools, sessions, and devices.
cq serve # Start the MCP bridge. That's it.CQ uses a relay server for NAT traversal. Traffic between your machines is end-to-end encrypted. Works behind corporate firewalls, WSL2, and dynamic IPs with zero configuration.
curl -fsSL https://raw.githubusercontent.com/PlayIdea-Lab/cq/main/install.sh | sh
cq init # Login + project setup (one-time)
cq claude # Start buildingUpdate anytime: cq update
| AI Tool | Integration |
|---|---|
| Claude Code | Native MCP — full tool access |
| ChatGPT | OAuth 2.1 remote MCP proxy |
| Cursor | Remote MCP via mcp.pilab.kr |
| Gemini CLI | MCP-compatible connection |
| Free | Pro | Team | |
|---|---|---|---|
| Price | $0 | $5–10/mo | Contact us |
| Mode | solo | connected | full |
| Knowledge (AI self-capture) | Local SQLite | Cloud (pgvector) | Cloud + shared |
| Hub GPU jobs | — | 100 calls/mo | Unlimited |
| Relay (P2P) | — | Included | Included |
| Research Loop | — | Included | Included |
| Team knowledge base | — | — | Included |
| Component | Description |
|---|---|
| Go MCP Server | 275+ tools (core + Hub + conditional), Registry-based |
| Knowledge | FTS5 + pgvector (OpenAI 1536d) + 3-way RRF + auto-distill |
| Hub | Distributed job queue, DAG engine, artifact store, cron, watchdog |
| Session | Auto-summarize via LLM, context injection on startup |
| Research Loop | Autonomous ML experiment cycle (plan→train→evaluate→iterate) |
| Paper Mode | Structured paper/document learning with knowledge DB integration |
| 42 Skills | Claude Code slash commands (/plan, /run, /finish, /pi, /paper, etc.) |
Architecture
┌──────────────────┐ ┌────────────────────────────┐
│ Local (Thin Agent)│ JWT │ Cloud (Supabase) │
│ │◄───────►│ │
│ Hands: │ │ Brain: │
│ ├ Files / Git │ │ ├ Tasks (Postgres) │
│ ├ Build / Test │ │ ├ Knowledge (pgvector) │
│ ├ LSP analysis │ │ ├ LLM Proxy (Edge Fn) │
│ └ MCP bridge │ │ ├ Quality Gates │
│ │ │ └ Hub (distributed jobs) │
│ Service (cq serve)│ WSS │ │
│ ├ Relay ─────────┼────────►│ Relay (Fly.io) │
│ ├ EventBus │ │ └ NAT traversal │
│ └ Token refresh │ │ │
└──────────────────┘ │ External Brain (CF Worker) │
│ ├ OAuth 2.1 MCP proxy │
Any AI (ChatGPT, ── MCP ──►│ ├ Knowledge record/search │
Claude, Gemini) │ └ Session summary │
└────────────────────────────┘
solo: Everything local (SQLite + your API key)
connected: Brain in cloud + relay (login + serve)
full: Connected + GPU workers + research loop
cd c4-core && make install # Build + install
cd c4-core && go build ./... && go test -p 1 ./... # Go tests
uv run pytest tests/ # Python tests
cq doctor # Health checkDocumentation | Installation | Quick Start | Architecture
Personal Study & Research License (Non-Commercial). See LICENSE.md. Copyright (c) 2026 PlayIdeaLab.