← Back to landing
Strategy A
Strategy B
Strategy C — recommended
Setup guide
Local vs Cloud vs Hybrid
Pick your install strategy based on the computer you actually have — not the one you wish you had.
ClickSorcery runs Jerry through OpenClaw. You choose where the AI brain lives: on your hardware (Ollama), in vendor clouds (direct APIs), or through a smart router (OpenRouter / DeepInfra). Telegram and OpenClaw software stay free — you only pay when thinking happens in someone else's cloud.
Three architectures
| Trait | Strategy A — Local (Ollama) | Strategy B — Direct cloud APIs | Strategy C — Hybrid (OpenRouter / DeepInfra) |
|---|---|---|---|
| Computer needed | Strong workstation — VRAM matters | Normal laptop (8–16 GB RAM is enough) | Normal laptop — compute lives in the cloud |
| Fixed monthly cost | €0 API (electricity only) | High & unpredictable — each vendor bills separately | Very low — pay per token you actually use |
| Data privacy | 100% on your machine | Enterprise contracts possible, but data crosses to cloud regions | Medium — traffic passes brokers; not for raw banking data |
| Biggest upside | No internet dependency for inference | Original Claude / OpenAI environments | Best price flexibility on the market |
| When to choose | 16 GB+ RAM or dedicated GPU, privacy-first | 8 GB laptop, need top-tier single vendor | Default sweet spot for most users |
Local computer + local models
You have a real AI workstation and want zero API bills and full privacy.
| Model | Hardware | Good | Avoid when |
|---|---|---|---|
| Qwen3.5:4b or 9b | 16 GB RAM laptop or 6 GB+ VRAM GPU | Very fast, low memory — good for routine sub-agent tasks | Struggles with very hard logic and long code debugging |
| Gemma4:26b-moe (or similar MoE) | 16–24 GB VRAM (RTX 3090/4080) or Mac 36 GB+ RAM | MoE architecture — smart and fast for its size | Heavy RAM/VRAM — can freeze a normal office PC |
| DeepSeek-R1:32b or 70b | AI workstation — dual RTX 4090 or Mac Studio 64–128 GB | World-class reasoning and code — €0 per message at home | Huge model files (40 GB+), slow load without top hardware |
Normal PC + direct APIs (Claude, OpenAI)
You need corporate-grade vendors, official channels, and absolute top models — on any laptop.
| Model | Hardware | Good | Avoid when |
|---|---|---|---|
| Claude Sonnet (Anthropic API) | Any PC — even an old 8 GB laptop | Excellent coding and long-context work in Cursor | Expensive — long chats burn balance fast (real users: ~€5 / ~10 questions + research) |
| OpenAI o1 / o3-mini | Any PC | Strong math and step-by-step reasoning | Closed stack, data to OpenAI, separate billing |
| GPT-4o-mini | Any PC | Fast, relatively cheap API baseline | Less capable than Sonnet or o1 for hard tasks |
Normal PC + OpenRouter / DeepInfra
Sweet spot: one API key, many models, penny-scale daily use.
| Model | Hardware | Good | Avoid when |
|---|---|---|---|
| DeepSeek-R1 (671B via router) | Any PC — cloud runs the model | Top-tier reasoning for cents — great demo of agent power | Slow replies while it “thinks” |
| DeepSeek-V3 or Llama-3.1-70B | Any PC | Best price/speed — chat all day for a few cents | Not offline; data leaves your machine |
| Claude Sonnet (via OpenRouter) | Any PC | Call Sonnet through same wallet as cheap models | Background agents can drain balance if unchecked |
What to pick in Magic Installer
| Under 16 GB RAM, no GPU | Cloud Commander | Do not choose Local Supercomputer — Ollama often never answers |
| 16 GB RAM, no GPU | Local possible (slow) or Cloud Commander | qwen2.5:7b class — OK for light Jerry + Telegram |
| 8–16 GB VRAM GPU | Local Supercomputer | Jerry + several agents, €0 API |
| 24 GB+ VRAM | Local Supercomputer (full company) | Optional cloud API for hardest tasks only |
Practical launch formula (hybrid + optional local)
- →Daily chat & Telegram: OpenRouter or DeepInfra (DeepSeek-V3, Llama-3.1-70B) — often a few euros per month.
- →Stuck in Cursor on architecture: route one session through Claude Sonnet via OpenRouter.
- →When you buy a strong PC later: run Ollama locally for routine agents; keep cloud for critical tasks only.