CLI · AI agents

Build AR scenes with the AI agent you already use

One command links a folder to an ARClip project. Claude Code, Codex or Cursor then builds the scene with the same tools as our editor — and checks its own work before it tells you it is done.

Terminal
npm i -g arclip
arclip login
arclip link <projectId>
arclip dev
claude
arclip verify && arclip render -o shot.png
arclip publish

No Node? One command

A standalone build installs with one line, the same on Windows, macOS and Linux — Astra Linux and RED OS included.

macOS and Linux
curl -fsSL https://staging.arclip.design/install.sh | sh
Windows · PowerShell
irm https://staging.arclip.design/install.ps1 | iex
Homebrew · macOS and Linux
brew install webar-studio/tap/arclip
All releases and downloads →

What you get

  • An agent that looks before it reports

    arclip verify catches broken references and scenes nobody can see, arclip render draws the scene on the server without an open editor, arclip playtest runs the scripts and presses every control. The agent sees the result instead of guessing.

  • Your scene on your phone, live

    arclip dev prints a QR code. Scan it and the draft opens in the phone's browser — no login, no publishing. The phone offers the new version after every edit, and the terminal re-checks the scene as it changes.

  • Your agent, your subscription

    Anything that runs terminal commands or speaks MCP works. arclip link registers the ARClip MCP server for the folder, pinned to one project — and your token never lands in a project file.

  • Live with your team

    The scene stays on the platform, so the agent's edits appear in your teammates' editors as they happen, and theirs reach the agent. Nothing to export, nothing to merge.

What arclip link sets up

Run it in any folder — a new one or your existing repository — or start with arclip clone, which also brings the project's scripts and UI cards down as files. Configs you already have are merged, not overwritten.

.arclip/project.json
Which project this folder belongs to. Safe to commit.
CLAUDE.md
Instructions for the agent: build, verify, render, playtest, then report — and never publish unless you asked. Kept if you already have one.
.mcp.json
The ARClip MCP server, started through the CLI and pinned to this project.
.claude/settings.json
Lets the agent read our docs and run the read-only checks without asking. Publishing always asks.

Commands

arclip --help lists them all, and every command has its own --help. The result goes to stdout, diagnostics to stderr.

Accountloginlogoutwhoami
Projectprojectsclonelinkpullpushdiffstatus
Developdevdev --watch
Checkverifyplaytesttestrenderrender --real
Publishpublishopenexportserve
Agentscalltoolsmcp serve

Scripts and CI

--json makes every result machine-readable, and a failed check exits with code 1. In CI, put a token and a project in the environment — no login needed.

CI
export ARCLIP_TOKEN=arclip_…
export ARCLIP_PROJECT=<projectId>
arclip verify --json
arclip render --view top -o top.png

Honest limits

  • Scripts and UI cards live in the folder as files (arclip pull / push); the scene itself stays on the platform, where people edit it together, and is not pulled as files yet.
  • arclip render is a fast server-side preview: layout, scale, colours and textures are real, scripts and effects are not. For the exact look, arclip render --real draws the scene with the published runtime in headless Chrome on your machine — 3D and surface scenes; ones that start from a camera need a device.
  • arclip playtest and arclip test play the scene in the editor, so the project has to be open in a browser tab — including in CI.
  • What an agent may do follows your plan: on Free it can read and check projects, while building, AI generation and publishing come with paid plans. AI generation spends your team's credits, as it does in the editor.
  • arclip dev opens the draft in the phone's browser, not yet in the ARClip app. A running AR session offers to reload rather than swapping the scene under you.

Questions

Is the CLI free?

Yes, on every plan. What an agent can change through it follows your plan's permissions, and AI generation spends credits exactly as it does in the editor.

Which agents does it work with?

Claude Code, Codex, Cursor and any other tool that runs terminal commands or connects to MCP servers. arclip mcp serve bridges any MCP client to your project.

Where is my token stored?

In your user config, readable only by you, and it is only ever sent to the server that issued it. Project files never contain it. You can revoke it at any time in your dashboard under API & MCP.

Can I use it in CI?

Yes. Create a token in your dashboard, pass it as ARCLIP_TOKEN together with ARCLIP_PROJECT, and use arclip verify --json as a check: it exits with code 1 when the scene has errors.

Do I still need the editor?

The editor is where people see and shape the scene together; the CLI is how agents and scripts join them. Both work on the same live project.

Can we host a project ourselves?

Yes. arclip export writes one folder — index.html, the scene as project.json, every file it uses and the player — that runs from any static host: your own server, S3, an intranet. Fonts are included; only fallback glyphs for scripts they lack (CJK, Arabic) still come from the ARClip CDN. Assets from the shared library keep their licence terms.

Give your agent an AR scene to build

Create a project, link a folder and ask your agent for a first draft. You can take over in the editor at any moment.

Start building — free