ML AI Code
Describe it once. A Frontend engineer sets the design, then ML AI Code writes, tests and ships the software around it. Every result is checked before it is called done.
See it work
No mockups. Each clip is the shipping app fixing the same bug: a checkout that charges twice when someone taps Pay twice.
On your own computer. It plans, asks before it edits, runs the check and shows the diff.
Plans before it builds
In Guided mode it reads the repository and proposes a short task list. Untick anything you do not want, then press Start building.
Asks before it edits
Each step arrives as a card that says what changes, why it is safe and how to undo it. Approve once, approve the rest of the task, or reject.
Checks its own work
Edits, commands and checks are grouped as they happen, so you can open any step. The answer says what the check confirmed, and a check that did not run is reported as not run, not as passed.
Runs in parallel
Type a follow-up during a run and choose Run in parallel. It starts in its own chat with its own machine, and both show live in the sidebar.
Review every change
Changes shows exactly what moved since the last checkpoint. Review it, run a security scan, export a patch, push it, or restore the checkpoint.
Works where you are
One account, one set of projects and chats. Start on your phone, review on your laptop, finish in the terminal.
Any browser
Each project gets its own isolated cloud computer with previews, a terminal and diffs side by side.
Open the workspaceMac and Windows
The same workspace as a native app. Choose This computer and the agent works in your folder, with files kept on your machine.
Download for desktopiPhone app and mobile web
Home, Library, Needs you and live runs, built for one hand. Approve a step from wherever you are.
Get the phone appYour terminal
Start tasks, approve diffs with one key and pull changes into your local repository.
curl -fsSL https://code.ml.ai/cli/install.sh | shDesign, led by Frontend
In Auto, a Frontend engineer owns the look of everything with a screen. It writes a short design brief, sets one visual system, uses real copy and imagery, then checks the result in a browser at phone and desktop sizes. It is part of every task, at no extra cost.
Tokens for colour, type and spacing are set once and used on every screen, so the product reads as one thing.
Sixteen by nine, premium layouts. Change words, images and order on the right, present full screen, export PDF or PowerPoint.
Hero and section images are made by the built-in image tool in one art direction. Icons, charts and patterns stay crisp in code.
Included with every task. No switch to turn on, no extra credits.
The problem
Then you hand one a real task in a real repository, and the same six failures show up in tool after tool. They are rarely about how clever the model is. They are about everything around it.
Asked to fix one bug, it rewrites three files nobody mentioned, and calls half a job finished.
Long tasks outgrow its memory. Your original request and your corrections are the first things to fall out.
One bad edit leaves a file that no longer parses, and there is no clean way back.
“All tests pass”, when no tests ran, a failure was hidden in a summary, or the code changed after the check.
A line of instructions is all that stands between the agent, your keys, a force push and production.
Close the laptop or lose the connection, and the work, and everything you decided, is gone.
The answer
Every one of those failures is an engineering problem, so we engineered the answer. Each system is ordinary, inspectable code with its own tests, sitting between your request and the result.
It drifts from what you asked.
Before the first edit, the request becomes a written contract: what to deliver, which files are in scope and which checks must pass. Progress is derived from that contract, and a task cannot finish while a deliverable is open.
It loses the thread.
Long tasks outgrow any working memory. Older tool output is summarised at a fixed threshold while the original request, every correction and the most recent steps stay word for word. Nothing is deleted: the full history can be read back.
It breaks what already worked.
Every change is checked as it is written. An edit that breaks a file which parsed a moment ago is refused, a checkpoint is taken before each step, and parallel helpers work in their own copies so nothing collides.
It says done when it is not.
"Done" is treated as a claim. The run records which commands ran, how many tests they covered and whether the code changed afterwards. Claims without matching evidence get one correction round, then are reported as unproven.
Only a prompt fences it in.
Each command is classified before it runs. A short list is always refused, a repository file sets the rules for that project, and a team can only tighten those rules, never loosen them.
It forgets, and it stops.
Work survives interruptions. Actions with side effects are journaled before they start, decisions and open questions carry to the next run, and specialists can be resumed with what they already found.
Measured, not claimed
Public tasks from widely used open-source projects, graded outside the agent on a clean copy. We publish first attempts and later repairs as separate numbers, and we say plainly what the evidence does not show.
Development evaluations on selected public tasks, not an official score. Method and limits.
Vision
Mission
Build the engineering that makes coding agents dependable: they stay in scope, check their own work, and tell you plainly what they could not prove.
How it works
It reproduces the problem, edits the code, runs your tests and starts a live preview, while you focus on the decisions.
Reads the repository and writes a task list you can follow.
Changes only what the plan names, with a checkpoint first.
Executes commands and tests on its own machine.
Opens a live preview and a diff you can review or turn into a PR.
Pricing
Every plan has the full agent. Monthly or annual, save 15% or more on annual. Plans differ by credits, CPU hours, PR reviews and team features.
Exploring ML AI Code
Indie developers and side projects
Professional engineers shipping daily
Engineering teams of 2 to 50
ML.ai Forge
Forge is built and hosted by ML.ai for long agentic coding tasks. On Team, run it all day on every seat without spending a credit.
Real GitHub issues fixed in real repositoriesIndependent run
Multi-step work in a live terminalModel card
Graduate-level science reasoningIndependent run
Fresh competitive programming problemsIndependent run
Forge is a pruned build of an open-weight model. Scores shown are the published scores of that base model: SWE-bench, GPQA Diamond and LiveCodeBench from independent runs, Terminal-Bench 2.1 from its model card. Our own Forge evaluations are in progress, and Forge may score differently.
Enterprise: custom pricing with SSO, SCIM, VPC or on-prem, BYOK and a dedicated engineer. Contact sales. Credit top-ups (Starter and Pro): $5 for 40 credits, $10 for 80 credits, $15 for 120 credits. Frontier credits, CPU hours and PR review packs roll over and never expire. Pay by card, Apple Pay or Google Pay through Stripe. Prices in USD, excluding tax.
Questions
A coding agent in your browser. It takes a task, writes a plan, edits code on its own cloud machine, runs your tests and previews the app, then reports exactly what changed.
Not in Guided mode. It shows the task list first and you pick what to do. Commands wait for your approval, or you can approve the rest of a task in one tap. Auto mode is there when you want it.
Pixis and ML.ai Forge, our own coding model, run on the credits included in your plan: 10 on Free, 75 on Starter, 200 on Pro and unlimited on Team, so Forge has no cap on Team. Frontier models run on frontier credits at the provider’s list price plus a small plan fee (15% on Free down to 8% on Team). Cloud build time is counted in CPU hours.
A task in progress finishes cleanly with a summary; new tasks wait until you add capacity. Starter and Pro can top up credits at $0.125 each, up to $15 a cycle, then upgrading is the better deal. Frontier credits, CPU hours and PR reviews come in packs that roll over and never expire. Auto-recharge is off unless you turn it on.
Shared projects, assigning and claiming tasks, handoffs, continuing a teammate’s task, a task board, roles with review and release controls, free viewers, unlimited credits, pooled add-ons, SSO, audit log and per-member limits. No seat minimum.
Each project gets its own isolated cloud computer with the common languages, databases, Docker and cloud CLIs preinstalled. You can also run the agent on your own laptop, where files never leave it.
Yes. Manage or cancel in the Stripe billing portal from Settings. Your plan stays active until the end of the period.
Write yours. Your first machine is ready in seconds.
