theLabs is an applied-AI lab that ships working software, not slideware. So this page is not a pitch. It is the working: three products, three derivations, one admission at a time.
Claim under test · the local-LLM default
“It fits — the file is smaller than your VRAM.”
Kayon refuses that arithmetic. It adds up what actually occupies the card, for your GPU, at your context length, and lets you expand any verdict to see the working.
And where the arithmetic cannot be done honestly (SSM, linear, hybrid attention), Kayon returns a refusal instead of inventing a number:
UNVERIFIED_ARCH
Worked example · 8B model · Q4_K_M · 8 GB card · 8k context
Claim under test · what Cempala promises
“Ask one agent for something and it pulls in another. Same machine, no second window, no copy-paste, no cloud service in the middle.”
Worked example · an illustrative delegated request, not a captured transcript
Claim under test · what Dalang claims to be
“A coordinated pool of models, delivered as one model. Point your existing SDK at it and keep your code.”
Run the proof yourself
Ordered by what you can do right now, not by what we are proudest of.
An honest local-LLM workstation.
Every quant of every model gets a verdict computed for your GPU at your context length, and you can expand any one of them to see the arithmetic it came from.
Not yet code-signed, so SmartScreen will warn. Code signing is on the roadmap.
One agent asks. The other just does it.
A local MCP server that lets Claude Code, Codex CLI, Antigravity and OpenCode hand work to each other in a single turn, instead of you carrying output between windows.
curl -fsSL https://raw.githubusercontent.com/thelabs-id/cempala/main/scripts/install.sh | bash
macOS / Linux. The Windows one-liner is on the product page. Learn more
A multi-agent system delivered as one model.
Point an existing OpenAI or Anthropic SDK at one endpoint and keep your code. Behind the single model string semar, a coordinated pool answers as one voice.
Pricing is published at general availability, not before.
An agent workspace where nothing leaves unseen.
Every agent-proposed change is a typed diff you approve or reject before it applies. Nothing here is built yet — this is the requirements document, not a product.
Nothing to install yet. The page describes what we're building against.
Five rules this site is built under. They bind the products too: the same honesty, enforced in code.
If you cannot run it, read it, or measure it, it does not go on the page. The film above is made of the products’ own vocabulary: verdicts, enforcement states, wire paths. Not invented figures.
Windows-only. Not a sandbox. Not SWE-bench. Invitation-only. The admissions are the credibility, not a tax on it.
Two products install in the next five minutes; one you request. The difference is the single most useful thing a visitor can learn here, so it is kept sharp.
A lab claiming “all the magic, none of the mystery” cannot ship a slow, heavy site without contradicting itself. The film is code, not video. Nothing on this page waits for a download.
No fabricated benchmarks, customers, testimonials, or logos. Where a number is an example, it says so on screen.
theLabs is an applied-AI lab in Indonesia. Research, engineering, design, and support under one roof.
The tools were built to prove themselves: Kayon’s arithmetic is visible, Cempala’s audit log is append-only, and Dalang’s benchmark harness was written before the thing it measures. The site holds itself to the same rule. No invented customers, no imported numbers, nothing you have to take on faith.
About the labKayon and Cempala are free, local, and need no account. Dalang is invitation-only: tell us what you are building and we will talk.