Local · Air-gapped · Yours
We install the whole pipeline — models, scripts, schedulers — directly on your computer. Nothing routes through the cloud, because there is no cloud.
Services
Remote
We join a screen-share session and build the automation with you watching — models pulled to your disk, scripts wired to your files, every step tested before we disconnect. You keep working access; we keep none.
from $400one-time setup, per automation
Offline
For work that can’t touch a network at all. You ship an encrypted SSD, we process it on an air-gapped machine, and the drive comes back with the results. No remote access, no copies retained.
from $650per drive, processed & returned
Tutoring
We teach you to run local AI yourself — picking and quantizing models, wiring prompt pipelines, scheduling jobs, and fixing things when they break — on your own hardware, at your pace.
$250per hour
Why local
Your files, prompts, and outputs stay on hardware you physically control. There’s no third party to trust, no account to breach, nothing for anyone to subpoena from a vendor.
The paid-service alternative: a hosted API or a rented GPU server processes your data on machines someone else runs. Every prompt and document crosses the network into their infrastructure, where it can pass through request logs, retention windows, and subprocessors you never see — and where a breach or a subpoena lands on them, with your data in scope. A zero-retention contract narrows that exposure but doesn’t remove the middleman.
You pay once to get set up. After that the compute is electricity you’re already buying — no subscription, no usage dashboard climbing in the background.
The paid-service alternative: hosted models bill per token; a rented accelerator bills per hour — roughly $1–$2+ for a capable GPU whether it’s working or idle. Costs scale with every job you add and never stop. A pipeline that classifies a few thousand documents a month can run into the hundreds monthly, indefinitely, for work a local machine does on power it’s already drawing.
It keeps running when the connection drops, when a provider rewrites its API, or when a cloud company sunsets the product. The automation is yours and it stays put.
Hardware
A model has to fit in memory alongside its context before it can run at all. That one constraint decides which models are realistic on your machine. The figures below assume 4-bit quantized models — the format most people run locally — with room left for a working context window.
| Memory | Model class that fits | What to expect |
|---|---|---|
| 8 GB | 3B – 4B | Short summaries, tagging, simple extraction. Reasoning is shaky — keep each task narrow and well-defined. |
| 16 GB | 7B – 8B | The practical floor for a general-purpose assistant. Handles most drafting, classification, and question-answering over your own notes. |
| 24 – 32 GB | 12B – 14B | A clear step up in reasoning and instruction-following, or an 8B model with a long context window. Comfortable for daily work. |
| 48 GB | 27B – 32B | Strong on nuanced writing, code, and multi-step tasks. Close to mainstream cloud quality for most jobs. |
| 64 GB | 70B (4-bit) | Where local output starts to rival premium cloud models, with roughly 40 GB of weights and headroom for a large context on top. |
| 128 GB + | 70B at full precision, or large mixture-of-experts models | Room to keep several models loaded at once and hold long-running context in memory — the basis for always-on agents. |
Recommended — what we build on
Memory is unified: the GPU can use nearly all of it. A Mac with 64 GB runs models that would otherwise need a data-center card, and one around $2,000 covers most real workloads.
Speed tracks memory bandwidth, so an M‑Pro, M‑Max, or Ultra chip is markedly faster than a base M chip at the same model size. This is the simplest, cheapest path to running large models locally.
Workable, but more expensive
System RAM and graphics memory are separate pools, and the model has to fit in the graphics card’s memory. Consumer cards top out near 24 GB, so anything past a ~32B model either doesn’t fit or spills over into system RAM, where it runs many times slower.
Running larger models here means buying a graphics card that costs $2,000 or more, using several of them together, or accepting slow processing on the CPU alone. We can still set it up — but expect to spend more on hardware, and more time getting it tuned, to reach the same result.
On the scoping call we look at what you already own, tell you plainly which tier you’re in, and lay out what it will and won’t do before you spend anything on hardware.
Built with this approach
A desktop app we built end to end on the local model. It connects straight to your mailbox over IMAP and SMTP; every AI step runs on a model on your own machine. No inbox contents are ever sent to a third-party service.
Case study
It reads a local cache of your inbox and does the repetitive triage work — sorting, summarizing, and drafting — while every send stays a deliberate human action. The same pattern (local model, cached data, tools that only read) is how we build most automations.
Unsigned builds for Apple Silicon and Windows. Needs a local Ollama install and two small models — setup steps are in the readme.
Sorts new mail into urgent, needs-reply, FYI, newsletter, or spam, and pulls out the concrete action items — so the queue is already ranked when you open it.
A two or three sentence summary of any thread, generated on demand and cached, so a long back-and-forth takes a glance instead of a scroll.
“What did Dana last send?” The assistant answers only from your cached inbox, through a fixed set of read-only tools — it never guesses and never reaches outside.
Replies are written for you and land in a Drafts tab. Nothing goes out over SMTP without an explicit, confirmed click — there is no auto-send anywhere in the app.
The Launcher
The same engine we install by hand, packaged as a one-download app. It scans your Mac, picks a local model sized to your hardware, and runs it through a bundled copy of Ollama — nothing to install first. Chat runs entirely on your machine, with a growing set of connected add-ons.
See the Launcher →Connections
The first add-on for the Launcher: it reads your inbox, drafts AI replies, and can auto-send for rules you explicitly turn on — all running on the same local model, no separate account or cloud service.
See Cloak Forge Email →Book a call
A 20-minute call to scope the work — what you run now, what hardware you have, and which of the three routes fits. No sales script.