Quick Start

GPU Anywhere, Anytime, Anything — from install to first result in 2 minutes.

1. Install

curl -fsSL https://raw.githubusercontent.com/PlayIdea-Lab/cq/main/install.sh | sh

Expected output:

Downloading cq v1.58 for darwin/arm64...
Installing to /usr/local/bin/cq
cq installed successfully.

Run: cq init

Verify:

cq --version
# cq v1.58 (darwin/arm64)

Linux / WSL2: The same one-liner works. cq is a single static binary with no runtime dependencies.

---

2. Initialize

cq init

cq init is the single command that handles everything: login, project setup, MCP registration, and hooks. After GitHub OAuth in your browser:

Logged in as changmin@pilab.co.kr

Setting up project...
  [OK] .c4/ directory created
  [OK] .mcp.json written (275+ tools registered)
  [OK] CLAUDE.md installed
  [OK] Git hooks configured

CQ is ready. Start your AI with: cq claude

Free tier: You can use CQ locally with your own API keys without logging in. See Tiers for differences.

---

3. Start Your AI

Launch your preferred AI directly from CQ:

cq claude    # Claude Code
cq cursor    # Cursor
cq codex     # OpenAI Codex
cq gemini    # Google Gemini

CQ registers 275+ MCP tools automatically. No manual .mcp.json editing required — cq init already handled it.

Run a quick health check:

cq doctor

[OK] Binary: /usr/local/bin/cq v1.58
[OK] .c4 dir: /your/project/.c4
[OK] .mcp.json: registered (275 tools)
[OK] CLAUDE.md: present
[OK] Hooks: pre-commit, post-commit configured
[OK] Python sidecar: running (pid 12483)
[OK] Hub: connected (relay.pilab.kr)
[OK] Supabase: reachable

---

4. First Commands

Search your knowledge base

Knowledge search is an MCP tool — ask your AI directly:

"Search my knowledge base for GPU setup"

The AI calls cq_knowledge_search automatically and injects results into context. On a fresh install this returns nothing — your knowledge base grows as you work. After a few sessions:

Found 2 records:
- GPU worker setup (2026-03-28) — score 0.91
  "Connect RTX 4090 via cq serve start, latency 8ms."
- CUDA environment (2026-03-15) — score 0.87
  "Always run nvidia-smi before submitting jobs. Driver 525+ required."

Start the relay

cq serve start

Relay: relay.pilab.kr:443
Latency: 11ms
E2E encryption: yes (ECDH + AES-256-GCM)
Your relay address: usr-changmin-a3f8.relay.pilab.kr

Ask Claude to do something

Inside Claude Code (or any MCP-connected AI), CQ tools are now available. Try:

"Search my knowledge base for anything about authentication patterns"

Claude calls cq_knowledge_search automatically and injects results into context.

Found 3 records:
- JWT validation: always verify signature, never trust alg:none (2026-03-10)
- Auth middleware pattern: use context.WithValue for user propagation (2026-03-18)
- OAuth flow: state param required to prevent CSRF (2026-03-22)

Record something

"Save a note: prefer table-driven tests in Go, always name subtests descriptively"

Claude calls cq_knowledge_record. Next session, this preference surfaces automatically.

---

5. What's Happening Behind the Scenes

Every session, CQ:

1. Captures decisions, corrections, and experiment results
2. Extracts preferences from what you corrected or repeated
3. Promotes patterns that appear 3+ times into hints, 5+ times into permanent rules
4. Syncs across all your AI tools — learn in Claude, use in ChatGPT

By session 5, CQ knows how you work without being told. By session 30, it codes like you.

See Knowledge Loop for the full picture.

---

What Next?

I want to...	Go to
Connect a GPU server for remote training	Worker Setup
Understand Free / Pro / Team differences	Tiers
See how knowledge accumulates over time	Knowledge Loop
Set up E2E encrypted relay (NAT traversal)	Relay
Connect ChatGPT to my CQ workspace	Remote Brain
Run GPU experiments autonomously	Worker Guide

---

Troubleshooting

cq doctor    # Diagnose issues automatically

Symptom	Fix
"MCP server not found"	Run cq init again or check binary path in .mcp.json
Login fails / browser doesn't open	Run cq init --no-browser and paste the URL manually
Relay connection refused	Check firewall: outbound TCP 443 must be open. Then run cq serve start
cq doctor shows auth error after update	Run cq init again — token may have expired
macOS code signing error	Install via the curl one-liner, not cp from another machine