Open Design turns the coding agent you already run into a local design studio through one od command

od is a daemon binary, and the desktop app sits on top of it

The root package.json maps a single binary, od, to ./apps/daemon/bin/od.mjs. That daemon lives in the apps/ workspace beside packages/, skills/, design-systems/, design-templates/, plugins/ and a figma-plugin/ directory. The root package is marked private and pinned to [email protected], with workspace:* dependencies such as @open-design/daemon and @open-design/components, so this is a pnpm monorepo you run as a product, not a library you import. The package description spells out the loop in one line: detect the code-agent CLI you already have installed, run design skills and design systems through it, and stream the resulting artifacts into a sandboxed preview. Generation happens in your agent. Open Design supplies the skills, the templates, the brand assets and the preview surface, which is why the project calls itself local-first.

MCP install targets and the DeepSeek Harness native runtime are two different integrations

Open Design connects to agents in two ways. Agents that consume OD get skills, a CLI and an MCP server. Agents that OD launches directly get a native runtime adapter. The compatibility table lists 19 agents and platforms, and 18 of them use the same command shape with a short target name:

od mcp install claude

The other targets include codex, cursor, opencode, cline, kimi, kiro, trae, raven, reasonix, vibe and hermes. VS Code with GitHub Copilot and the GitHub Copilot CLI share one target, copilot. DeepSeek Harness is the exception. You install the official dsh CLI first, then either pick it inside Open Design or run:

od agent setup deepseek-harness

That command installs or repairs OD's connection component. Structured streaming, model discovery, cancellation and session resume are named for this native runtime; the MCP rows in the table carry a supported status and a setup command, with no comparable feature list attached.

od mcp install can show its changes before writing them

Each MCP install edits another tool's configuration, so the dry-run flag matters more than it looks. Add --print to see what would be written without applying it:

od mcp install <agent> --print

The matching --uninstall flag removes the integration again, and the full list of targets comes from:

od mcp install --help

For a developer who keeps hand-tuned MCP settings in Cursor, Cline or Codex, the sensible order is print first, read the diff, then install. Because one target name, copilot, covers both the VS Code extension and the standalone Copilot CLI, the print output is also the quickest way to see which of the two configurations od is going to touch on your machine before it does so.

Studio splits a project into Prototype, Slides, Image and HyperFrames

Inside a project, Studio offers four modes. Prototype builds and refines interactive web experiences with your agent next to a live preview. Slides creates decks, lets you review slides and speaker notes, and exports when ready. Image generates visual assets from the conversation, then lets you inspect, download or open the result. HyperFrames creates motion graphics with the agent and exports them to MP4. Around Studio sit the core pages: Home starts from a brief or an example, All projects revisits design systems, media and prototypes, Design Systems lets you inspect brand assets and refine them with an agent, and Plugins lists skills you can launch with a Try it button. The 0.26.0 release is titled Share, Review, Refine, matching the stated workflow of sharing a prototype for team feedback and letting the agent refine it in the same project.

Claude Desktop on Linux gets no automatic MCP configuration

The Claude Desktop row in the compatibility table carries a footnote: automatic MCP configuration is currently supported on macOS and Windows only. Running od mcp install claude-desktop on a Linux workstation therefore does not give you the one-command setup that the table implies for every other row. The consequence is practical. A Linux user who works through Claude Desktop rather than Claude Code has to wire the MCP server into Claude Desktop by hand, and keep that hand-written entry in step with od as releases land. If your team is mixed, the Linux members are the ones who will hit this first. Using Claude Code on Linux avoids the footnote, since its row, od mcp install claude, has no platform restriction attached.

Without an agent CLI you are routed to BYOK or the paid Cloud

Everything in the local workflow assumes an installed coding agent for od to detect. If you have none, the README points to a BYOK proxy endpoint and to OpenDesign Cloud as the alternatives. The code is Apache-2.0, but the hosted path has a price: the OpenDesign Go plan currently costs $8 for the first month, with up to $120 in monthly credits for eligible Design Plan models, usable across more than ten models including GPT-6 Luna, DeepSeek V4.1 Flash, GLM-5.3 Flash-X and MiMo V2.6 Flash. Personal plans also include API access for Codex, Claude Code, DSH and OpenCode. The consequence for a reader hoping for a free, self-contained design tool is that output quality and running cost both come from the agent or model you bring; Open Design itself does not ship one.

Ollama and other local model runners have no row in the compatibility table

The compatibility table names hosted-model agents and agent CLIs: Claude Code, Codex CLI, Cursor, OpenCode, Kimi CLI, Mistral Vibe CLI, Hermes Agent and others. Ollama is not one of the rows, so there is no od mcp install ollama target and no native runtime adapter for it. People searching for a way to run Open Design on a local model will not find a one-command path in that table. The workable routes are indirect: an agent CLI from the list that you have configured against your own model endpoint, or the BYOK proxy. Neither route is covered by the supported status marks in the table, so expect to debug the connection yourself rather than rely on a documented setup.

Three releases in one week, and a package.json version that trails the tags

Open Design 0.25.0, titled OpenDesign Motion, was tagged on October 15, 2026. 0.26.0 followed on October 20 and 0.26.1, titled Intelligent Router, on October 22. The most recent push to the repository was on November 2, 2026, so the project is under active development. The root package.json still reads version 0.25.1, behind the open-design-v0.26.1 tag. For anyone scripting around od or pinning a build for a team, the version string in the source tree is not a reliable indicator of which release you have. Check the release tag you downloaded, read CHANGELOG.md before upgrading, and expect the command surface to keep moving at this pace. The repository also carries a .node-version file and a mise.toml, which pin the toolchain for anyone building from source.

Editorial conclusion

Open Design fits a team that already pays for a coding agent such as Claude Code, Codex or Cursor and wants that agent producing prototypes, decks and MP4 motion pieces against its own design system, with review and refinement inside one project. It fits poorly if you need a Linux Claude Desktop setup without hand editing, a local model through Ollama, or a version that stays still for a month. Before adopting it, run od mcp install with --print for your agent, read the BYOK proxy section of the README if you have no agent CLI, and check which release tag you are actually installing.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes