OpenDesign: 92.8K stars, a local design bridge for Claude Code, Cursor, and Codex (2026)

OpenDesign is a skill-like agent workspace worth a sandbox trial when your coding agent needs to produce a real visual artifact rather than a text-only answer. The project has 92.8K GitHub stars, 10.7K forks, and a latest release dated August 25, 2026; a representative OpenDesign skill page shows 2.3K installs. Those numbers show attention and reach. They are not independent proof that every generated artifact is production-ready. 1 2 3

The useful first move is small: install OpenDesign into one agent, ask for one local landing-page artifact, inspect the files, and keep credentials and remote systems out of the test. OpenDesign currently runs as a local-first workspace that connects Claude Code, Codex, Cursor, Cline, and many other agents to prototypes, dashboards, slides, images, and video exports. 4

What the workspace changes

A coding agent normally returns code in the conversation and leaves you to assemble the design process around it. OpenDesign adds a local daemon, an MCP server, plugins, templates, and a visual workspace. The documented flow runs from a brief to a plugin, design direction, design system, artifact, handoff, and memory. The output can be a project file or an export such as HTML, PDF, PPTX, MP4, ZIP, or Markdown. 1

The agent still does the reasoning and terminal work. OpenDesign supplies a route for the agent to create and inspect visual deliverables. The README describes a filesystem-based preview for canonical project files and a sandboxed srcdoc iframe for artifacts assembled through a BYOK or plain-API route. 1

OpenDesign's official site names 21+ coding agents and lists adapters for Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot CLI, OpenCode, OpenClaw, Antigravity, Cline, Kimi, Kiro, Pi, Mistral Vibe, Hermes, and others. The README also gives explicit od mcp install commands for supported hosts. A host listed by the project still needs a local agent executable that OpenDesign can discover. 1 4

The local-first model has a clear cost boundary. The official site says that OpenDesign sends your prompt and skill context to the provider you bring, while provider API charges remain yours. The agent can read local files and execute terminal commands, so the important security decision happens before the first real repository or credential enters the workflow. 4

Install one host

Choose the host you already use. OpenDesign's current release is 0.21.0, marked Latest on August 25, 2026. The release notes mention repaired agent detection, a one-line installer that completes dependency resolution, Kimi ACP startup fixes, and a configurable stdio MCP idle lifetime. 2

The desktop route is the shortest path. Download the macOS or Windows app from the official download page, then let the app detect the agent executable on your PATH. The project lists Linux through an optional AppImage release lane. Source runs require Node around 24 and pnpm 10.33.x. 1 4

After the od command is available, install the MCP bridge for one agent:

od mcp install claude

Use the matching identifier for another host:

od mcp install codex
od mcp install cursor
od mcp install cline

The project also documents a hosted shell equivalent:

curl -fsSL https://open-design.ai/install.sh | sh -s claude

The first command configures Claude Code. The other commands target Codex CLI, Cursor, and Cline. The README lists Cline as a supported target, while the official site describes adapter-based compatibility rather than a single universal runtime. Start with the host you can verify locally. 1 4

Preview the change before applying it when you are working on a shared machine:

od mcp install --print

Remove the bridge with:

od mcp install --uninstall

The project also exposes commands for listing projects, reading files, listing installed skills, and managing plugins:

od project list --json
od files list <project-id> --json
od skills list --json
od plugin search "landing page"
od plugin list

On macOS, the name od can resolve to Apple's /usr/bin/od, the octal-dump utility. The README tells desktop users to use the absolute-path MCP snippet under Settings → MCP server when that collision appears. OpenDesign's WSL2 path has its own setup guide. 1

The default security posture gives the first test a useful boundary: the MCP server is read-only, the daemon binds to 127.0.0.1, and connector credentials and live-artifact preview routes remain loopback-only. LAN access requires an explicit bind host and allowed-origin configuration. Provider URLs resolving to private or link-local ranges are blocked by default. 1

OpenDesign uses Apache-2.0 for the project. Bundled third-party skills retain their own licenses, so check the individual skill or plugin before redistributing a generated package. 1

First test: one local landing page

The first test should verify four things: the host can discover OpenDesign, the agent can create a visual artifact, the output lands in a known folder, and the agent can explain what it changed. Use a temporary project directory and keep the request below bounded:

Use OpenDesign in this temporary project to create a one-page landing page for a local developer tool.

Constraints:
- Work only inside this project directory.
- Create real HTML, CSS, and any required local assets.
- Use a restrained charcoal, warm amber, and coral palette.
- Give the page one clear hero, three feature blocks, and one call to action.
- Keep the layout responsive at mobile and desktop widths.
- Do not use production credentials, remote deployment, or unrelated files.

Before editing:
1. Confirm that the OpenDesign MCP connection is available.
2. State which files you will create or change.
3. Ask for clarification if the brief requires an external asset or an unavailable model provider.

After editing:
1. List every changed file.
2. Start the local preview if the installed host supports it.
3. Report the preview command and the output folder.
4. Stop after the local artifact and a short verification report.

A successful first run leaves an inspectable folder and a repeatable preview command. Open the HTML and CSS yourself. Check that the page has a working responsive layout, that the call to action points somewhere intentional, and that the generated asset list contains only files you expected.

The OpenDesign catalog also exposes focused entries such as frontend-dev and web-artifacts-builder. Each page shows 2.3K installs and points readers to the upstream project or, for web-artifacts-builder, Anthropic's source repository. Those entries are useful as discovery surfaces; the upstream skill remains the place to retrieve its original scripts, references, and assets. 3 5

Second test: hand off the artifact to code

Use the second test when the first page renders and you want to see whether the design-to-code handoff fits your existing workflow. Keep the task inside the same temporary project:

Use the existing OpenDesign landing-page artifact in this temporary project.

Prepare it for handoff to the coding agent:
- preserve the current visual direction and content hierarchy;
- identify the design tokens for color, spacing, type scale, and border radius;
- write a short DESIGN.md that records those tokens and the component rules;
- keep the implementation in plain HTML and CSS unless the project already uses a framework;
- point out any generated asset that needs a license check;
- run a local render or syntax check if the available tools support it.

Show the proposed file changes first. Apply them only inside this project. End with:
1. the files changed;
2. one command to rerun the preview;
3. two visual decisions that a human designer should review;
4. any unresolved dependency or runtime blocker.

OpenDesign's README describes this handoff as real HTML/CSS moving into Cursor, Codex, or Claude Code, with a DESIGN.md that can carry confirmed palettes, fonts, and artifacts into later sessions. The handoff is a workflow promise from the project documentation; your own render and file review decide whether the result fits your codebase. 1

The second test also tells you when to stop. A generated page that looks good in a preview still needs accessibility checks, content review, asset licensing, and ordinary code review before deployment. The agent can prepare those checks; the test itself does not complete them.

What users report

One recent open issue raises a local/cloud ownership concern. On August 7, 2026, GitHub user 52216108 asked whether signing into a cloud account caused local projects to appear as cloud projects, made the projects disappear after sign-out, and prevented conversion back to local ownership. The same issue also complains about a payment prompt around DeepSeek. The issue remains open under the enhancement label, and the page shows no maintainer response or linked pull request. The report is a user's question, so the project page leaves the behavior unresolved. 6

A separate Codex report shows a different kind of friction. On May 1, 2026, user coballester reported that chat input caused codex exec to exit with an unexpected - argument. The issue was closed after PR #342 changed the Codex spawn path so prompts arrived through stdin alone. The page also points to later related fixes, which makes agent-version compatibility a live part of the setup rather than a box to check once. 7

The adoption picture is broad but unevenly measured. GitHub shows 92.8K stars and 10.7K forks for the repository, while the skills directory shows 2.3K installs for representative entries such as frontend-dev and web-artifacts-builder. The directory also displays security-audit labels for those entries. These counts and labels describe public attention and the directory's screening surfaces; they do not replace a review of the exact plugin, script, dependency, and generated files you plan to run. 1 3 5

Use it now or keep it sandboxed

Try it now when your work matches all four conditions:

You want an agent to create a visual artifact, a design system, or a design-to-code handoff.
You can run the first task in a temporary local project.
Your host executable is installed and discoverable by OpenDesign.
You can inspect every changed file, dependency, asset, and command before sharing access.

Keep it in a sandbox when any of these conditions apply:

The workflow touches production credentials, a customer repository, or a remote deployment target.
You are signing into the cloud path while your local-project ownership behavior remains unclear.
Your machine resolves the wrong od executable, or the agent integration fails during discovery.
The artifact depends on a private model endpoint, a hardware-specific runtime, or a third-party asset whose license you have not checked.
You need a weekly star-velocity claim or an independent quality score. The current evidence shows strong attention, fresh release activity, and directory reach; it does not provide a verified seven-day star delta or a neutral benchmark of your exact workflow.

The smallest useful action today is the local landing-page prompt. It tests the connection, the artifact path, the file boundary, and the preview loop before OpenDesign sees a real product repository.

Quick reference

Field	Pick details
Pick	`nexu-io/open-design`: a local-first, agent-native design workspace with MCP, plugins, and real artifact exports. 1
Best for	Landing pages, prototypes, dashboards, slides, design systems, and design-to-code handoffs. 1
Hosts	Claude Code, Codex, Cursor, Cline, and other named adapters in the project's host list. 1
Install	`od mcp install claude`, or replace `claude` with `codex`, `cursor`, or `cline`. 1
Prerequisites	A supported agent executable; source runs require Node around 24 and pnpm 10.33.x. 1
Current release	`0.21.0`, marked Latest on August 25, 2026. 2
Adoption signal	92.8K GitHub stars, 10.7K forks, and 2.3K installs on representative skills.sh entries. 1 3
License	Apache-2.0 for OpenDesign; bundled skills can carry their own licenses. 1
First test	A local landing page with an explicit file boundary, preview command, and stop condition.
Main caveat	Local/cloud project ownership remains an open user question, and agent integration has had version-specific fixes. 6 7