The Launcher

Cloak Forge Launcher

A local-AI dashboard for your own Mac — no cloud, no subscription, one download.

What it is

Cloak Forge Launcher is a terminal app that turns a fresh Mac into a working local-AI setup in one step. Open it, and it scans your hardware, works out the biggest language model your machine can actually run well, and gets a local model server (Ollama) running — bundled in, nothing to install first. From there you chat with the model entirely on your own device. Nothing you type leaves the machine.

Cloak Forge Launcher terminal dashboard, showing the Connections section with the Email module installed

The dashboard — hardware, models, and connected add-ons, all in one screen.

What it does

  • Hardware-matched model picks. Scans your CPU/GPU and unified memory, then recommends a model sized to fit — from a fast 1.7B model on modest hardware up to 235B-class models on a high-end workstation, always the largest size that will actually run well rather than just the largest that downloads.
  • Download and remove models freely. Pull any size at any time with visible live progress, and remove one later to free disk space — re-download it again whenever you want. Nothing is one-way.
  • Local, on-device chat. Talk to whichever model you have active, with optional file, shell, and web-search tools it can call — each one permissioned to ask every time, auto-approve, or stay fully read-only, your choice.
  • Connections. A growing set of add-on modules that plug into the same local model instead of shipping their own separate AI — more on this below.

The models it can run

All 4-bit quantized, all Apache‑2.0 licensed. It always defaults to the largest general-purpose size your hardware can run well; the vision and coding sizes are opt-in — listed and downloadable, but never chosen for you automatically.

SizeDownloadWhat it's for
1.7B 1.4 GB Fast and light on almost any Mac. Fine for a single, simple tool call at a time; not for multi-step reasoning.
1.8B (vision, opt-in) 1.8 GB The lightest way to send it an image — basic description and detection, no tool-calling.
8B 5.2 GB A real step up in reasoning. Reliable for one tool call at a time; still shaky chaining several together.
7B (vision, opt-in) 6.0 GB Meaningfully better at reading documents, charts, and screenshots than the smaller vision size.
14B 9.3 GB A dependable everyday default — handles both single- and multi-step tool use with reasonable judgment.
30B (coding, opt-in) 18 GB Tuned for code and agentic tool use, with a 262K context window for large codebases. Stronger than the general 32B size at exactly this, weaker at open-ended writing.
32B 20 GB Strong general reasoning and dependable multi-step tool use — chaining tools, recovering from a bad result, nuanced instructions.
~70B-class (MoE) 28 GB A mixture-of-experts model giving roughly 70B-class quality at a fraction of the active compute. Chats well; can't call tools.
235B-A22B (MoE) 142 GB Frontier-class reasoning with full tool-calling. Realistic only on a very high-end unified-memory or multi-GPU workstation.

Connections

Add-on modules that plug into the same local model instead of shipping their own — one AI setup on your machine, a growing set of tools that share it.

Available now

Cloak Forge Email

Reads your inbox, drafts AI replies, and can auto-send for rules you turn on yourself — running on the exact model you already have loaded, no separate account or cloud service.

Why local

No API bill that scales with usage, no account, and no dependence on a provider staying up or staying priced the way it is today. The trade-off is honest: a local model is not going to out-think a frontier hosted one. What you get instead is a setup that is entirely yours — private by construction, working offline, and still yours in five years regardless of what any AI company does.

Cloak Forge Launcher

  • macOS · Apple Silicon
  • bundled Ollama — no prior install
  • hardware-matched model pick
  • local chat, on-device
  • permissioned file / shell / web tools
  • Connections — add-on modules, same local model

Unsigned build — macOS asks you to clear it on first open (one Terminal command, covered in the readme). Apple Silicon only for now.

← Back to the main site