Why we built an MCP server for design edits

Copy as prompt copies a Markdown specification that you can paste into your coding tool. MCP offers a different handoff: a configured client can read the session directly. This article explains the design rationale, not a version-by-version release chronology.

When Anthropic released Model Context Protocol in late 2024, it was obvious in retrospect what the right surface was. MCP is a standard for letting agents call tools. A design edit is a tool call. Match made.

We could have built a bespoke WebSocket protocol and an SDK per client. We chose MCP because it provides a shared tool boundary and supports both local stdio and remote HTTP transports, while still requiring client-specific configuration and verification.

The current tool catalogue is documented on the MCP page. Those tools do not themselves grant filesystem or general network access; a coding client may have separate capabilities and repository permissions outside Design Mode.

Local mode uses a stdio MCP server on your machine. Design Mode MCP traffic stays on localhost, latency is low, and the agent must run on the same machine. The trade-off is a steeper setup for people who are not comfortable with terminal commands.

Cloud mode exposes a Streamable HTTP endpoint at mcp.designmode.app and uses an authenticated event stream for the extension. Consult the canonical privacy disclosure for relay handling and retention limits. A client-specific configuration and bearer token are required; no local companion process is needed.

Self-hosted mode is the same Cloud relay code (packages/mcp-cloud) deployed on infrastructure you operate. For teams who want Cloud ergonomics but their own infra.

Before MCP, sending edits felt like a tax. With MCP, the hand-off can become the natural end of a design iteration: make the visual change, mark the session ready, then ask the connected client to read it and implement the source diff in a repository it can access.

MCP gives Design Mode a standard boundary for tools and prompts, but each client still has its own configuration, transport and authentication rules. The integration must be verified client by client.