Security & open models
Local AI, official Ollama sources, and your data
ClickSorcery runs local open-weights models through Ollama on your machine. Download apps and models only from official channels — not random forums, unknown GitHub forks, or third-party file hosts.
Official sources only
Use the official Ollama site and Ollama’s model library. That reduces the risk of tampered installers, bundled malware, or modified weights from untrusted mirrors.
- Ollama — official website
Install Ollama only from ollama.com. Check the browser address bar before downloading.
- Ollama model library (official)
Browse and pull models with official names (e.g. ollama run qwen2.5-coder). Downloads come from Ollama’s registry (registry.ollama.ai).
Which local models work well with ClickSorcery?
You choose the models; ClickSorcery routes agents to Ollama when Brain is set to local. Smaller models fit everyday chat; larger ones need more RAM/VRAM.
Qwen2.5 / Qwen2.5-Coder (7B–32B, up to 70B on strong hardware)
Strong everyday reasoning, multilingual use, and coding. A common default on 16 GB PCs is qwen2.5:7b; 32B+ when you have 24–32 GB+ unified memory or VRAM.
DeepSeek-R1 (distilled or 32B/70B where hardware allows)
Reasoning-heavy tasks — math, architecture, multi-step plans. Heavier than 7B chat models; plan RAM and pull time accordingly.
GLM-4, Llama 3.1 / 3.3 (tool-calling and multi-agent flows)
Useful when agents call tools and structured workflows; verify the exact tag in the official library before pulling.
Local open-weights vs cloud LLMs
Both are valid in ClickSorcery (local Ollama or your cloud API keys). This table is about trade-offs — not “one is always better”.
| Topic | Local (Ollama) | Cloud (API) |
|---|---|---|
| Data privacy | Prompts stay on your device when using local models. | Prompts go to the provider you connect; retention depends on their policy. |
| Offline | Works without internet once models are pulled. | Needs network and provider uptime. |
| Cost | Hardware + electricity; no per-token bill from ClickSorcery. | Pay per use or subscription to the API provider. |
| Peak capability | Up to 70B+ on workstation-class RAM/VRAM; limited by your machine. | Largest models in vendor datacenters; you pay for access. |
| Control | You pick models and versions from the official library. | Provider controls model behavior, limits, and API changes. |
Why local weights are not “mystery executables”
Weights are data, not arbitrary code
A pulled model (e.g. GGUF-based bundles via Ollama) is primarily numerical weights. It does not replace your OS or silently browse the web on its own.
Compute stays on your machine
Inference runs on your GPU, Apple Silicon, or CPU. ClickSorcery + Ollama do not send local prompts to ClickSorcery servers.
Air-gapped use
After install and pull, you can disconnect from the network and keep using local chat — useful for sensitive drafts (you still approve any real-world actions in the app).
Three checks for a safe local setup
1. Install from ollama.com
Download the Ollama app only from the official domain.
Verification: Browser shows https://ollama.com before you install.
2. Pull models via Ollama CLI/library
Use official commands (e.g. ollama run qwen2.5:7b) — not random .gguf files from forums.
Verification: Terminal shows pulls from Ollama’s registry (registry.ollama.ai).
3. Keep the API local
Ollama listens on localhost:11434 by default. Do not expose that port to the public internet without auth (Tailscale or similar if you need remote access).
Verification: Works at http://127.0.0.1:11434 on your PC; not open to the world without a deliberate tunnel.
Hardware vs model size (quick guide)
Parameter count (7B, 32B, 70B) needs RAM or VRAM. Unified memory on Apple Silicon counts toward the same pool.
⚠️ 8–16 GB — Basics & trials
Models: 3B–8B (e.g. qwen2.5:3b, qwen2.5:7b, llama3.1:8b)
Fine for light chat; tight for heavy agents, long context, or deep reasoning.
🟢 24–32 GB — Serious daily local use
Models: 14B–32B (e.g. qwen2.5-coder:32b, deepseek-r1:32b)
Recommended minimum for strong coding and agent workflows on one machine.
🔵 64–128 GB — 70B class
Models: 70B models (e.g. qwen2.5:70b, llama3.3:70b)
Workstation / Mac Studio territory — competes with premium cloud models for many tasks.
🟣 128 GB+ — Multi-agent / enterprise desks
Models: 70B+ and multiple models loaded over time
Run several local agents or large context with fewer compromises.
Cheat sheet
| RAM / VRAM | Typical model size | ClickSorcery fit |
|---|---|---|
| 8–16 GB | 3B–8B | Trial & light chat |
| 24–32 GB | 14B–32B | Strong daily local work |
| 64–128 GB | 70B | Pro / code-heavy |
| 192 GB+ | 70B+ / multi | Maximum local alternative to cloud |