.lcl — an AI workbench that runs on your machine

click anywhere for sound

AI on your own machine — nothing leaves it unless you say so

Download .lcl for Windowslatest signed release

This is the app's own opening screen. Download it to keep everything — or press Just chat to run a small model right here, in your browser.

↓ what the full app does

Lite session

not loaded — weights fetch once, then it is all local

⋮⋮Files

A real lite workspace, stored in this browser. The model reads these files when you ask about them. The full app links real folders on your disk — every change shown, every change revertable.

⋮⋮Tasks

nothing running

⋮⋮Activity

⋮⋮Permissions

networkthis page only

filesbrowser storage only

scriptsnever

hardwarenever

The full app is deny-by-default across 60+ tools — the dangerous ones show the exact action and ask first, every time.

this chat sits on a full workbench — see what it does

the full app

What you just saw, plus everything it exists for

An AI workbench that runs on your machine, with the network switched off. Everything is deny-by-default — this list is a permission map, not a brag sheet.

Offline is the default state

Unplug the network and it keeps working: local llama.cpp inference, on-device OCR, speech-to-text, image generation. The online parts are the ones you switch on.

60+ tools, every one gated

Read and edit files, run scripts, query CSVs with SQL, extract PDFs — and the dangerous ones show the exact action and ask first, every time.

A reference library that cites

~64 public-domain volumes indexed offline: ask an engineering question, get the document and page. One click downloads any missing source — or all of them.

It acts on real hardware

Identify boards on your bench, read their boot logs, back up flash, compile and upload firmware — each with the exact command shown before it runs.

Four places to think

This machine, a machine you own, an API vendor, or a rented GPU — chosen per conversation, never hidden, with a rented box never dressed as yours.

An error log that answers "why"

Every handler failure, rejected promise, and refused download leaves a JSON line in a local log — debugging starts with reading it, not screenshotting an error prompt.

what it is actually for

Things you can just ask it

Each of these is one sentence typed into the same chat you just tried. The difference is that the full app is allowed to act — after asking you first.

Go through last night's logs and tell me what broke.

Link the folder once; it lists, reads and searches the real files, and answers with the failing lines quoted.

file tools · linked workspace

Pull every table, image and form out of this 300-page PDF.

The extractor takes the whole document — born-digital or scanned — and hands the pieces back as files.

PDF extraction · OCR

What board is on COM4, and what is it running?

It identifies the silicon, reads the boot log, and can back up the flash before anything is touched.

hardware · serial · ask-first

What does the standard say about fillet weld sizing?

The offline library answers with the document and page cited — no internet involved.

reference library · citations

Merge these six CSVs and give me totals by month.

It queries the files in place with SQL — no import step, no spreadsheet wrangling.

SQL over CSV

Draw this system as a clean diagram.

Described in text, rendered on your machine — diagrams via Graphviz, images via local Stable Diffusion.

diagrams · image generation

how updates work

Signed one-click patches, from this repository

The app you download tonight patches itself from the same public repo you can read — and refuses anything that does not prove itself.

Cut and signed

Every release passes the full gate — thousands of checks plus a driven-UI harness — then the build is signed with an offline Ed25519 key. The public half ships inside the app.

Detected and verified

Your install polls this repo's releases. A newer official build downloads with its progress on the button, is hash-matched against the signed manifest, and rollbacks are refused — an unsigned or tampered file never launches.

Installed on your click

The app waits until the installer has genuinely started before it closes, and says so honestly if the Windows prompt was dismissed. Nothing changes without you.

	Minimum	Comfortable	To run everything
RAM in the machine	8 GB	16 GB	32 GB
Disk for models	6 GB	25 GB	60 GB
CPU	any x64	8 cores	16+ cores
GPU	not required	optional	optional
Network	never required	never required	never required

honest comparison

Lite vs full

The page you are on is the taste; the download is the workbench. Everything in the full column is permission-gated — none of it runs without your click.

	This page (lite)	The full app
Thinking
Local inference	WebGPU — 0.5–1.5B models	llama.cpp — a 4B daily driver, up to a 9B flagship, CPU or GPU
Where it can think	this browser only	your machine · a machine you own · an API vendor · a rented GPU — chosen per conversation, never hidden
Works fully offline	after the model caches	yes — offline is the default state
Multi-step work	single answers	plans, runs tool chains, audits its own output, concludes
A model that fits	fixed small models	the load planner checks your RAM first — a model that will not fit is refused with the shortfall stated, never a frozen machine
Concurrency	one chat	local chats queue honestly — RAM is finite; API and node conversations run wide
Workspace
Sessions	saved in this browser	on your disk, every one resumable
Files	notes in browser storage, read by the model	real folders — listed, read, edited; every change shown, every change revertable
Terminal	—	a real terminal beside the chat — human keystrokes only, never the model's
Tools
Tool count	4 demo stubs that explain themselves	60+, permission-gated
Scripts & automation	—	writes and runs them — the exact command shown before it runs
Data	—	SQL over your CSVs, no import step
PDFs	—	extracts everything — text, scans, tables, images, forms, annotations
Vision & images	—	OCR on screenshots and scans · image generation on your own silicon · diagrams rendered via Graphviz
Speech	—	local dictation — rolling transcription, audio never leaves the machine
Knowledge
Reference library	—	~64 public-domain volumes indexed offline — answers cite the document and page
Your own documents	—	add them; they join the same cited index
Filling the shelf	—	a missing volume downloads in one click — or the whole shelf at once
Hardware
Real devices	—	identifies boards, reads boot logs, backs up flash, compiles and uploads firmware — ask-first, every step
Trust
Permissions	read-only demo	deny-by-default — dangerous tools name the exact action and wait for your click
Cost	$0.00, always	$0.00 local — API spend, when you choose it, tracked to the cent
Updates	—	signed one-click patches from this repo — unsigned, tampered or older builds refused
Logs	—	every failure and refusal written to a local error log you can read
Audit trail	—	every tool call and refusal recorded in a local audit log you can read
Account / API key	none	none
Telemetry	none	none
		Get the full appfree · MIT · Windows

questions with real answers

FAQ

Does anything I type here leave my machine?

No. The lite chat runs the model in your browser via WebGPU — the weights download once from the model's public repository and cache locally; after that, generation happens on your GPU. This page has no analytics, no trackers, and no server of its own. Your sessions and files live in your browser's storage, on your machine.

Why does the download weigh ~1.7 GB?

It brings its own everything: llama.cpp, stable-diffusion.cpp, whisper.cpp, Tesseract, ffmpeg, qpdf, ImageMagick, SQLite, Graphviz, a starter model, and the offline reference index. No Python, no Docker, no account — it answers the moment it installs.

Is my data used to train anything?

No. There is no cloud side. Local models run locally; if you deliberately link an API vendor, that conversation goes to that vendor under their terms — and the app names the destination on a card before anything leaves.

How do updates work?

The app checks this repository's releases. A newer official build downloads with visible progress, is verified against an Ed25519-signed manifest baked at build time, and installs only on your click. Unsigned, tampered, or older builds are refused outright.

Can I read the source?

All of it — this page, the app, the engine, the tests, and the release tooling live in the public repository, MIT licensed. The test suites are the honest spec: thousands of checks, many quoting the exact failure that forced them.

Mac or Linux?

Windows today. A Mac build is planned; the lite page runs anywhere WebGPU does.