MCP server

MCP — the Model Context Protocol — is the open protocol that AI assistants use to call external tools. Display.dev exposes an MCP server with publish, share, rename, find, get, export, and delete tools; once your assistant is configured, you can ask it to publish artifacts in plain language.

What is MCP?

If you've used Claude, Cursor, or Codex with extra tools, you've used MCP. The host (the assistant) connects to an MCP server, discovers the tools it offers, and calls them with structured arguments. Display.dev's server exposes the artifact surface — publish, share, rename, find, get, export, delete — as MCP tool calls. Branding and account-level admin operations stay CLI-only.

Three install paths

Display.dev's MCP server is reachable three ways. Pick whichever fits how you already work:

• Custom connector (recommended for Claude Desktop / Claude.ai / ChatGPT) — paste /v1/mcp into the host's "Add custom connector" field. The host runs an OAuth flow against a browser tab opens, you sign in with your work email, pick the organization, and approve. No config file, no API key. Revoke from the dashboard at any time.
• Remote endpoint with an API key — your assistant connects to /v1/mcp over HTTPS with Authorization: Bearer sk_live_…. Best for headless setups and hosts that don't support OAuth, or when you want a long-lived credential pinned to a specific machine.
• Local CLI subprocess — your assistant spawns dsp mcp as a subprocess and talks to it over stdin/stdout. Requires the @displaydev/cli CLI installed locally; auth is the session from dsp login, no key to manage.

All three produce identical artifacts; switch between them at any time.

Configure your assistant

Each host has a different config file. The host-specific guides cover all three install paths above:

• Claude Desktop — supports all three (Add custom connector, mcp-remote with API key, or local CLI)
• ChatGPT — Custom connector with auto-discovered OAuth
• Claude Code — .mcp.json (project) or ~/.claude.json (user)
• Cursor — .cursor/mcp.json or ~/.cursor/mcp.json
• Codex — ~/.codex/config.toml

For Claude.ai (web), add Display.dev as a custom connector under Settings → Connectors → Add custom connector. Paste /v1/mcp as the URL and follow the OAuth prompts in the popup. The web client doesn't support local-stdio.

For other MCP-aware tooling (custom clients, hosted agent platforms), the remote endpoint accepts either an OAuth token (via the dynamic-client-registration flow above) or Authorization: Bearer sk_live_… with a static API key — pick whichever fits the host's auth model.

Tools exposed

• publish — publish an artifact from a path or inline content.
• share — change visibility or sharedWith.
• rename — change an artifact's display name.
• find — list or search the organization (filters: author, visibility, date; cursor pagination).
• get — fetch artifact metadata.
• export — fetch the source bytes (Markdown or HTML) of an artifact, capped at 1 MiB per call.
• delete — remove an artifact.

Each tool is a thin wrapper around the matching CLI subcommand. There's no extra surface, by design — composition happens in the assistant's prompt, not in the protocol.

Anonymous mode (no auth)

If dsp mcp is launched without auth (no dsp login, no DISPLAYDEV_API_KEY), the server exposes only the publish tool. Each call returns a public preview URL and a one-shot claim URL; the artifact has no owner until someone signs in and claims it. The remote endpoint always requires a credential — anonymous mode is local-CLI only. See Publish without an account for the full flow.