Open Design Quickstart: Run the AI Design Agent with Codex, Claude Code, Cursor, or OpenCode
A practical Open Design setup guide for developers who want to run the open-source AI design agent locally, connect coding agents, and generate the first prototype safely.
This Open Design quickstart is for developers who want to test the open-source AI design agent with their existing coding-agent workflow. The goal is not to connect every account on day one. The goal is to produce one safe artifact, inspect the output, and decide whether Open Design belongs in your design or product workflow.
Start with the Open Design profile, then keep the official Quickstart open while you install.
1. Pick the right install path
Open Design documents two main routes: a local development path and a Docker path. The local path currently expects Node.js 24 and pnpm through Corepack. The Docker path is better if you want a containerized service on a predictable port without managing the full Node toolchain.
| Path | Best for | Tradeoff |
|---|---|---|
| Local dev | Contributors and agent builders who want full source access | Requires Node 24, pnpm, and local debugging comfort |
| Docker | Teams that want a cleaner service-style run | Still needs token/env setup and Docker operations |
| Desktop release | Users testing the packaged app | Watch release notes and platform-specific bugs |
2. Run a minimal local test
The official local development path is roughly: clone the repo, enable Corepack, install dependencies, and run the web toolchain. The repository also includes desktop, daemon, web, skills, plugins, and design-system directories, so expect a large project rather than a tiny UI generator.
For Docker, the quickstart points to the deploy directory, an environment file, a generated token, and docker compose. That route is useful when you want persistent data and fewer local package-manager surprises.
3. Connect an agent through MCP
Open Design's README documents one-line MCP installation commands such as od mcp install codex, od mcp install claude, od mcp install cursor, and od mcp install opencode. This is the important product idea: Open Design is not trying to be one model. It is trying to be a design workflow that many agents can use.
If an agent CLI does not appear as installed, check the PATH used by the app or daemon. The official quickstart specifically calls out macOS GUI PATH differences, which can hide globally installed CLIs from desktop apps.
4. Use a harmless first prompt
Do not start by asking Open Design to redesign your real product dashboard with private metrics. Use a harmless prompt such as a fictional SaaS landing page, a sample mobile onboarding flow, or a synthetic KPI dashboard. The point is to evaluate structure, visual quality, export behavior, and how well the selected design system constrains the output.
A good first prompt includes the artifact type, target audience, required sections, visual direction, and export expectation. For example: create a one-page onboarding prototype for a fictional developer analytics product, use a restrained enterprise design system, include three states, and keep it exportable as a single HTML artifact.
5. Check the result like an engineer
After the first run, inspect the generated artifact. Does the layout fit the brief? Are components accessible? Does text overflow? Are images real or placeholders? Does the HTML export open outside the app? Can a coding agent continue from the artifact without rewriting the whole thing?
Also inspect operational behavior. Watch CPU usage, model costs, saved artifact location, logs, and any permissions requested by the agent. GitHub feedback already includes issues around AMR behavior and macOS CPU usage, so these are not theoretical checks.
6. Decide whether to expand
If the first artifact is useful, the next step is a real but low-risk workflow: a marketing page draft, a design-system exploration, an internal deck, or a prototype for a non-sensitive feature. Only then connect stronger credentials, private repositories, production brand assets, or recurring automations.