Klarcode
Made for Scientist, Engineer, and Entrepreneur. Any type of Builder and Innovator.
Klarcode is an AI agent and a full code editor in one Mac app. Tell it the job — a feature, a part, a simulation, a film. It works in the real tools, checks its own work before it says done, and hands you the finished file. When you want the keyboard, the editor is the same window.
Klarcode carries a guide for every model family that names how it tends to fail — and corrects it.
A check counts only if Klarcode ran it. Then a second model reads the report against your files.
CAD, simulation, games, film, spreadsheets — finished in the tools your field uses, and opened on your Mac.
⌘K 2 and you’re in the full editor while the agent keeps working. Select lines, ⌘K, and say the change.
Long runs don’t lose the plot. Your rules are re-stated every time the conversation is condensed.
Risky commands wait for you, Deny first. Choose Local only and the agent has no network at all.
The problem
Any good model can write code now. What decides whether the job gets done is everything around it — and that’s where today’s agents let you down.
The agent says it’s finished. No test, no check, no proof.
Switch to a cheaper or local model and the agent falls apart.
Your work lives in CAD, simulations, spreadsheets and video too.
Long tasks drift once the conversation gets too long.
The agent is in one window, the editor in another.
Your references, every message. Your outputs, buried up the chat.
Klarcode is the harness. Bring any model. It makes it do the work.
The method
Whichever model you pick, Klarcode wraps every turn in the same harness — so a fast, inexpensive model finishes work it would leave half done on its own.
How you work
With most coding agents, every handoff between the chat, the editor, the terminal and the browser is yours. In Klarcode they happen in one window.
Today, with a coding agent
Five apps. Every handoff is yours.
With Klarcode · one window
The workspace
Real screens from Klarcode. Everything you need around the agent — your references, what it made, the editor, the checks and the preview — in the same window.
Your brand sheet, the drawing, the spec — kept with the project. Switch References on and the next message carries them.
Real work
Klarcode does the work in the programs your field already uses, then hands you the file — a FreeCAD part, an Excel workbook, a Godot game, a PyMOL session — with a button that opens it where it belongs. Not a summary of the work. The work.
Every Klarcode window on this page is the real app, photographed on 21–22 October 2026 from tasks run on a MacBook Pro with a model connected by the tester’s own key. Where another program is shown, it was opened on the file Klarcode made.
Video and games
This is Klarcode’s own launch film, and Klarcode made it — 1920 × 1080, 20 seconds, mastered to −14.0 LUFS.
Any model
Most of the work on this page was done by fast models like DeepSeek V4.1 Flash and GLM 5.3 Flash. Tap a model to see its tasks.
The same seven everyday tasks, run inside Klarcode on 20 October 2026, each answer checked against the files. A test of Klarcode with each model, not a comparison with other tools.
Compared
| Feature | Chat assistant | AI code editor | Cloud coding agent | Klarcode |
|---|---|---|---|---|
| What you get back | Text to copy | Code changes | A pull request | The finished file, open in its app |
| Runs your desktop apps — CAD, Excel, Godot, Xcode | No | No | No — it works in their cloud | Yes, on your Mac |
| Checks its work before it says done | No | Varies | Runs tests | Your checks, plus its report read against the files |
| Any model, or one on your own machine | Its own | Varies | Its own | Any model, your key, or local |
| Your files stay on your machine | No | Partly | No | Yes — Local only keeps the agent offline |
| Take over and finish it yourself | Copy and paste | Yes | In another editor | Same window, ⌘K 2 |
Each column describes the typical tool of its kind in October 2026, not any one product.
Who it’s for
Scientists, engineers, developers, shop owners and creators use it for different work. Pick yours and see a real task, start to finish.
Scientists
A structural biologist needs a protein measured before a meeting. Instead of an afternoon of scripts, she asks once.
They ask
“Download ubiquitin from the Protein Data Bank, measure it, and open it in PyMOL.”
Klarcode
ResultRadius of gyration within 0.003 Å of an independent check.
Beyond tests
Tests are the start. Klarcode carries checks written for each kind of work, and runs them on the files it made before it says done.
Mechanical
Re-solves on a finer mesh and flags a stress singularity instead of reporting it.
Bracket: 581 / 583 / 580 MPa on three meshes
CAD
Volume, centre of mass and inertia computed from the part itself, not estimated.
Turbofan: 822,929 mm³, matched its report
Numerical methods
Claim second-order and Klarcode re-runs your solver to prove it — or fails it.
Euler measured 1.01, not the claimed 2: failed
Finance
Future data is changed three ways; past results must not move. Paper only — it can’t place orders.
Planted look-ahead leak: caught
Electronics
KiCad checks, ngspice simulation, and eye diagrams with interpolated timing.
555 timer: ERC 0, DRC 0, 0.986 Hz
Video
Length, frame rate, sample rate and loudness read from the finished file.
Launch film: 20.000 s, −14.0 LUFS
iOS
Privacy manifest, entitlements, icons and launch time checked from the built app.
Golden screenshots: a 4% colour shift caught
Games
What the report says is counted against the actual scene files.
Godot: 9 of 9 game checks passed
Each grey line is a run recorded in Klarcode’s own test notes, August–October 2026.
Safety
Klarcode plans, waits for your yes, and asks before anything risky — a command that changes your repository, an email, a payment, a deploy. You choose how much it asks, from every step to only what matters.
Security
Klarcode looks for leaked keys, unsafe commands and switched-off encryption, on your Mac. No AI reads your files to do it, and a clean result says only that these checks found nothing.
Deploy
Deploy runs only the command you set, and only after the project’s checks pass. A project that declares no checks cannot deploy unless you allow it.
The details
Pause, then type to resume with your words added.
A count it can’t measure shows a dash, never “+0”.
Permission you give today doesn’t quietly last forever.
Even in its most hands-off mode. Every time.
If an approval card disappears, the answer is no — never yes.
Klarcode reveals them in Finder. You decide to open them.
Type while it works: Send now, Edit or Cancel — or let it steer the run.
Your own keys and local models carry on. The conversation is intact.
Old failures are labelled, not blamed on this task — or hidden.
1,009 file types recognised, so results read at a glance.
Work done by a model on your Mac says so, and costs nothing.
Autocomplete runs on a model on your Mac. No cloud.
Privacy
The first thing Klarcode asks is how you want to work. Choose Local only and the agent has no network at all. Open Privacy and read every route that leaves your Mac — marked Closed, Open, or Unverifiable when it can’t prove it.
Freedom
Pick a model for each task, bring your own key from any OpenAI- or Anthropic-compatible provider, or run one on your Mac. Klarcode carries its own guide for each model family, and your own key is never slowed against what we sell. Then connect what you already use — GitHub, Notion, Linear, Figma, Blender, FreeCAD and 37 more.
Any model
DeepSeek, Qwen, GLM, Kimi, Gemini, GPT, MiniMax and models on your Mac. Klarcode carries its own guide for each family, so every one works the way it works best.
Freedom
Any OpenAI- or Anthropic-compatible provider, in three steps. Your key, your provider, your bill — and never slowed against what we sell.
Local models
Klarcode runs models on your own Mac with its built-in engine, and finds the ones you already have. Nothing leaves the machine, and no tokens are billed.
Offline
Choose it on the first screen: only models installed on this Mac, and no network for the agent at all.
Privacy
Open Privacy and read exactly what leaves your Mac. Telemetry: closed. The code index stays on this machine.
Your rules
Open, No training, or Strict — set per project. Every task keeps a record of which model saw your code and where that provider is based.
Yours to keep
Plain files in your own folders and plain git. If a plan lapses you lose the agent, not the project.
Control
A command that changes your repository waits for you, with Deny as the default answer. Emails, payments and deploys always ask.
Reviewable
Each file it touched, with its lines and a View changes that opens the diff. Nothing changes behind your back.
Honest
Every task ends with what was verified, in plain words. A check it could not run is named, never rounded up to done.
Make it yours
Add your own skills by folder or link, your own plugins, and any MCP server. Klarcode becomes the tool your work needs.
Ready check
Before a job starts it tells you which tools are ready, which are missing and why — not halfway through.
Getting started
The folder you already have — a repository, a CAD model, a set of spreadsheets. Nothing to import.
In plain words, the way you would brief a colleague. Klarcode plans it and asks before anything risky.
The finished file, the checks it ran, and every change it made, listed with its lines.
Built on the open-source base of VS Code: your keybindings, git and terminal are where you left them.
Pricing
Every plan is the full Klarcode: agent mode, manual mode, local models, your own key. What changes is how much runs at once. A plan renews monthly or annually until you cancel, and you can change it from your account whenever the work changes. A Founding 888 seat is the same price paid once, for 40 days (13 months on annual), and then it ends.
One developer who wants the full verified agent loop.
Rp480.000
Paid once, for 40 days
Rupiah only: QRIS, GoPay or virtual account
Several agents at once, with desktop control.
Rp1.580.000
Paid once, for 40 days
Rupiah only: QRIS, GoPay or virtual account
Specialist and hardware teams working at full scale.
Rp3.280.000
Paid once, for 40 days
Rupiah only: QRIS, GoPay or virtual account
A founding seat is the full product, bought once at the price shown, for a fixed term rather than a licence: it runs longer than the price normally buys — 40 days instead of 30 on a monthly term, 13 months instead of 12 on an annual one — and then it ends. The first 10 days run on your own model key, a local model or Klar Credits; the plan's included usage covers the rest of the term. Nothing renews and there is nothing to cancel. Each product has its own run of 888 numbered seats, assigned in payment order and never reused; one per person per product. Rupiah prices are the full amount; card payments may have tax added at checkout, depending on your country. Your files, git history and sessions stay yours in every state.
If Klarcode is not right for you, write to us within 3 days of paying — plan or founding seat — and tell us why. Once we have reviewed it, we refund it in full, as the refund policy sets out. A regular plan on a credit card can start with a 3-day trial; rupiah payments have no trial.
Files, repositories, git history and sessions survive every subscription state. A lapsed plan loses the agent, not the project.
Bring a provider key or a local model and it is never throttled, delayed or degraded relative to what we sell. Public promise, with a test behind it.
Every model's price will be published on a public rate card before launch, so you can see what each one costs against your usage.
Three ways you can check on this page. It plans before it edits and the plan is gated. It asks before it runs anything risky, with the risk read from the command itself. And a run ends in a verdict backed by named checks, with the ones it could not run listed instead of rounded up to green.
That is the case it is built for. Klarcode reads the repository first, so unfamiliar code is the normal starting point rather than the hard case.
It prepares the change and the pull request. A push to a shared branch is a higher-risk ask you approve. A deploy is refused by default when any check could not run.
Only what a request needs, to the model route your data policy allows, and never to train anything. Private mode routes only to providers verified not to train. A local model sends nothing at all.
You keep every file, repository, git history and session. What stops is the agent and the included usage. You cancel from your account, in the same place you subscribed. There is no call to book.
