Гайд 2026: Как бесплатно получить токены Gemini CLI, Claude Code, Copilot и развернуть на облачном Mac

Baseline snapshot. Proxy = нужен VPN/прокси для прямого доступа из РФ/СНГ без relay.

Stack math: SiliconFlow 20M + Alibaba 70M + Zhipu 20M = 110M tokens без подписки. Gemini CLI OAuth даёт до 1000 invocations/day — при avg 2K output tokens это theoretical ceiling ~2M tokens/day, на практике меньше за счёт Flash-Lite routing.

Install path (Node 18+): npm install -g @google/gemini-cli | brew install gemini-cli | zero-install: npx @google/gemini-cli. Auth: gemini → «Sign in with Google» → browser OAuth → token cached locally.

Runtime commands: /model, /model gemini-2.5-flash, /stats model. ToS constraint: OAuth token must not be proxied to third-party gateways — account ban risk. Enterprise/API-key path exempt from 2026-06-18 Antigravity cutover for personal OAuth users per Google blog.

OpenAI Codex CLI — 83K+ stars, kernel-level sandbox, headless CI capable. Install: npm install -g @openai/codex | brew install --cask codex. Auth path A: ChatGPT OAuth (codex → browser). Path B: CN relay without VPN.

~/.codex/config.toml example:

openai_base_url = "https://api.siliconflow.cn/v1"
model = "deepseek-ai/DeepSeek-V3"
sandbox_mode = "workspace-write"
web_search = "disabled"
approval_policy = "on-request"

Env: export OPENAI_API_KEY="sk-...". Health check: codex doctor. One-shot: codex "fix rustc error E0382 in src/main.rs".

No permanent free tier. Install vectors: curl -fsSL https://claude.ai/install.sh | bash, brew install --cask claude-code, npm install -g @anthropic-ai/claude-code, Windows: winget install Anthropic.ClaudeCode.

Relay via ~/.claude/settings.json:

"env": { "ANTHROPIC_BASE_URL": "https://api.siliconflow.cn/v1", "ANTHROPIC_API_KEY": "sk-xxxx" }

Or shell: export ANTHROPIC_BASE_URL=... export ANTHROPIC_API_KEY=... claude. Model switch: /model | claude --model claude-sonnet-4-6. Critical: never run /init on monorepo — single invocation can burn 100K–500K tokens indexing entire tree.

OpenCode install: curl -fsSL https://opencode.ai/install | bash | brew install anomalyco/tap/opencode | Docker: docker run -it --rm ghcr.io/anomalyco/opencode. Provider wiring: TUI /connect or ~/.config/opencode/config.json.

OpenClaw: curl -fsSL https://openclaw.ai/install.sh | bash, openclaw onboard --install-daemon. Gemini OAuth reuse: openclaw models auth login --provider google-gemini-cli --set-default. Claude token: claude setup-token → openclaw models auth paste-token --provider anthropic. Free model discovery: openclaw models scan. Production deploy: OpenClaw на Mac Mini M4.

Copilot Free: GitHub Settings → Copilot → Enable. Limits: 2000 completions/mo, 50 premium requests/mo. Copilot Student: GitHub Education → auto Pro parity, 300 premium/mo. Note: since 2026-04-20 new Copilot Pro signups may be paused — check UI state. OSS maintainers: github-copilot/signup/orgs.

Cursor Hobby: download from cursor.com, 2000 Tab completions + 50 slow premium/mo, no CC. Student Pro: cursor.com/students. Usage dashboard: app.cursor.sh/account/usage. Agent workflows: Copilot Coding Agent runbook.

Universal env pattern for Codex/OpenCode/Claude relay:

export OPENAI_API_KEY="sk-..."
export OPENAI_BASE_URL="https://api.siliconflow.cn/v1"

OpenRouter free models complement CN stack — routing policy in LLM trends June 2026.

• Rule 1: ban /init on monorepos — Claude Code/Codex will ingest entire tree.
• Rule 2: single-file scope per prompt, not «refactor entire codebase».
• Rule 3: route trivial tasks to gemini-2.5-flash-lite (1000 RPD) before Pro.
• Rule 4: set 80% quota alerts on SiliconFlow/Alibaba dashboards.
• Rule 5: local Ollama for lint/format — ollama pull qwen3:8b, zero API burn.
• Rule 6: compress system prompts in CLI configs — 5–15% overhead reduction per call.
• Rule 7: version SKILL.md for recurring prompts — reduces multi-CLI drift.
• Rule 8: openclaw models scan for OpenRouter permanent free models.
• Rule 9: multi-key rotation when hitting HTTP 429 rate limits.
• Rule 10: track vendor promo campaigns — CN platforms run seasonal credit drops.

Rent vs buy: 3–6 month token experiments → hourly cloud Mac from страница цен, then lock spec on заказ. Dimension by agent parallelism and Docker sandbox count, not weekly model hype.

Free tiers solve model access. Dedicated Apple Silicon solves OAuth session persistence, launchd KeepAlive, and tenant isolation. Aligns with console provisioning runbook.

1. 01
   Provision instance: NUKCLOUD консоль — pick region, 16 GB for light CLI OAuth, 32 GB+ for OpenClaw/Hermes + Docker. SSH key, verify disk quota for OAuth token cache and agent logs.
2. 02
   SSH baseline: ssh user@your-cloud-mac, xcode-select --install if needed, brew install git node python@3.12. Ping Google OAuth endpoints, GitHub, and your chosen API BASE_URLs.
3. 03
   Register free API keys: SiliconFlow, Groq, Gemini AI Studio. Write to ~/.zshrc: export GEMINI_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY. Enable 80% quota alerts on provider dashboards.
4. 04
   Install CLI stack: npm install -g @google/gemini-cli @openai/codex, OpenCode via curl, OpenClaw via openclaw onboard. Complete Gemini OAuth in browser; run codex doctor; OpenCode /connect.
5. 05
   launchd 7×24 gateway: ~/Library/LaunchAgents/com.team.freeagent.plist with RunAtLoad + KeepAlive for OpenClaw/Hermes. Centralize prompts in SKILL.md — see Hermes install guide.
6. 06
   Monthly quota retro: Gemini /stats model, Cursor/Copilot usage pages, SiliconFlow console. If API spend consistently exceeds high-memory Mac rental — optimize routing; else lock RAM tier. Pricing: цены, order: заказ.

Shared minute-pool macOS VPS resets long OAuth sessions and kills launchd gateways — fatal for 12h agent loops with thousands of tool calls. NUKCLOUD dedicated cloud Mac nodes provide auditable uptime for Gemini OAuth cache and parallel CLI processes.