Quick start
No setup. Open the workspace, describe what you want and the first machine starts on its own.
Open the workspace
Sign up, or try it as a guest. A blank project is ready straight away.
Describe the work
Type it or say it. Attach files, screenshots or a GitHub repository.
Review the plan
Untick anything you do not want. The rest stays visible as cancelled.
Check the result
Read the diff, open the live preview, then keep it or restore a checkpoint.
How a task runs
Every task follows the same seven steps. Three of them are yours; the rest happen on the project’s machine and are recorded as they go.
- 1AskDescribe the work in plain words
- 2PlanIt reads the code and writes a task list
- 3ApprovePick the tasks you want done
- 4EditA checkpoint is saved before each step
- 5RunCommands and tests on its own machine
- 6VerifyEvidence is matched to each task
- 7ReportDiff, preview and what is unproven
A task cannot mark itself finished while a deliverable is open. If something could not be proven, the report says so instead of saying done.
Modes and effort
The mode decides how much the agent may do before asking. Switch at any time from the box where you type.
Plan
Drafts an approach and never touches your files. Use it to think a change through.
Brief
Reads the repository, proposes a task list and asks its open questions. Reply with changes as often as you like; once you approve, it builds the agreed list.
Guided
You approve the plan and the files in scope once. Edits inside that scope, tests and builds run on their own. Installs, network access, deletes and edits elsewhere ask first.
Auto
Runs builds without approvals. Deletes, history rewrites and force pushes still ask, and a few things are always refused.
| Effort | What it does | Good for |
|---|---|---|
| Low | Fast, light replies. | Questions and small edits |
| Medium | Thinks hard when planning and when something fails, moves quickly through routine steps. | Most work |
| High | Reasons through every step. | Tricky bugs and large changes |
Machines
Each project gets its own isolated cloud computer with common languages, databases, Docker and a browser already installed. Dependencies are set up once and redone only when your manifest files change. Each browser check opens a fresh browser profile, so saved state such as localStorage is only kept within one check.
2 CPUs · 4 GB
Scripts, quick fixes and questions.
4 CPUs · 8 GB
Everyday apps, tests and previews.
8 CPUs · 16 GB
Full-stack builds, Docker and databases.
Your laptop
Run the agent locally. Files never leave your computer.
Roles and skills
Give the agent a job title and it changes how it plans, what it checks and which skills it reaches for.
Engineering roles
Frontend, backend, data, mobile, security, release and more. Pick one with /role, or let the task decide.
Skills
Short how-to guides for common stacks. Only a one-line summary is loaded up front; the full guide is read when a task needs it. Add your own with /skills.
Put a short notes file in your repository and every task reads it first: conventions, commands, things not to touch.
Apps and secrets
Connect the tools your work already lives in. Sign in once and the agent can use them, with your approval, in any project.
One-tap sign-in
Search for an app, sign in on its own page and come straight back. On a phone it opens in a sheet you can close.
Repositories
Import a repository, keep it in sync both ways and open pull requests from a finished task.
Vault
Store keys once for your account. Commands receive them as environment variables; they are never pasted into code or shown in replies.
Actions in a connected app ask for approval even in Auto mode. Every write is scanned for secrets before it is saved or pushed.
Safety rules
Rules live in a file in your repository, so they are reviewed like any other change. A team can tighten them for everyone, never loosen them.
# .mlai/policy.yml
production:
branch: main
require_pr: true # no direct pushes to production
require_ci_green: true # every check must pass before promotion
min_approvals: 1
block_paths_without_review: ["infra/**", "payments/**"]
agent:
modes_allowed: [plan, brief, guided]
commands:
allow: ["npm test", "pytest*"] # run without asking
deny: ["terraform apply*"] # never run, in any mode
secrets:
scan: true
ignore_paths: ["tests/fixtures/**"]| Always refused for the agent | Why |
|---|---|
| Piping a download into a shell inside a workspace | Runs code nobody has read. The CLI install command is for you to run on your own computer; to read it first, download install.sh, open it, then run sh install.sh. |
| Publishing packages | Releases go through the release flow. |
| Force pushes | Rewrites history other people depend on. |
| Reading SSH keys | Credentials stay out of the conversation. |
| Production deploys from the terminal | Production goes through the deploy review only. |
Shipping
From a finished task to production, with a person in the loop where it matters.
Pull request
Open a draft from the task. It needs a passing check newer than the last change, unless you choose to publish without one.
Preview
Deploy a preview to Vercel, Netlify or Cloudflare and open it on any device.
Promote
Production needs a review. The card shows the checks, the diff and who may approve.
Roll back
Every deploy is listed. Go back to an earlier one in a tap.
Teams
Invite people by email and give each one a role. Shared projects show who started what, and a team audit log records approvals and deploys.
| Role | Can |
|---|---|
| Viewer | Reads runs, changes, deploys and logs. Cannot start or approve anything. |
| Contributor | Starts runs, approves plans and commands, edits files, opens pull requests. |
| Maintainer | Everything a contributor can, plus connected apps, preview deploys, direct pushes and repository rules. |
| Releaser | Everything a maintainer can, plus production deploys, rollbacks and releases. |
| Admin | Everything, plus members, invites, team rules and the team audit log. |
| Owner | Created the team. Everything an admin can, plus deleting the team. |
Apps and CLI
One account everywhere. Start on your phone, approve on the train, finish at your desk. See all downloads for what is available on each platform.
Web
The full workspace: projects, previews, diffs and a terminal side by side.
Mobile
Built for one hand. Voice input, approvals and previews on the go.
Command line
Run tasks from your own shell and sign in with your account.
curl -fsSL https://mlai-mini-code.vercel.app/cli/install.sh | shCommands
Type a slash in the message box for shortcuts.
/go | Skip the plan questions and build right away |
/new | Start a new project |
/plan | Switch to Plan mode |
/build | Switch to Guided mode |
/role | Choose an engineering role |
/diff | Open the changes |
/schedule | Schedule a recurring task, e.g. /schedule weekdays 9am run the tests |
/connect | Connect an app |
/mcp | Add a tool server |
/skills | Manage skills |
/agents | Open helper agents |
/browser | Open the live preview |
/import | Bring in an earlier session |
/publish | Publish a page, report or deck to a shareable link |
/settings | Open settings |
/doctor | Run a health check and fix problems |
/center | See every chat in one place |
/help | Open the command menu |