The Model Context Protocol (MCP) promises a universal standard for connecting AI models to external tools, but the implementation reality remains stubbornly fragmented. A new guide published on DEV.to by Usman Basheer dissects the exact steps required to connect a single remote MCP server—SVG Lab’s Design Engine—to five distinct AI clients: Claude, ChatGPT, Codex, Cursor, and OpenCode. The exercise highlights a critical friction point in the current LLM ecosystem: while the protocol is standardized, the client-side configuration interfaces are not. For Anthropic’s clients, the process diverges sharply between the web/desktop app and the command-line interface. In Claude.ai and Desktop, users must navigate to Customize, then Connectors, and manually add a custom connector named 'SVG Lab' with the URL https://svglab.app/mcp. OAuth fields are left empty before connecting. Conversely, Claude Code users must utilize the CLI command claude mcp add --transport http --scope user svglab https://svglab.app/mcp, followed by an in-terminal authentication flow via the /mcp slash command. OpenAI’s ChatGPT introduces a significant barrier to entry for developers testing MCP servers. Access requires 'Developer mode,' which is gated behind specific subscription tiers including Plus, Pro, Business, Enterprise, or Education. Users must enable this mode in Settings, then create a 'developer mode app' within the Plugins section, specifying the MCP server URL and OAuth authentication. This contrasts with simpler implementations elsewhere, suggesting OpenAI is still treating MCP access as a privileged developer feature rather than a standard utility. The developer-focused tools Codex, Cursor, and OpenCode offer more direct configuration paths but require manual file edits. Codex users can run codex mcp add svglab --url https://svglab.app/mcp or edit ~/.codex/config.toml directly. Cursor requires users to modify ~/.cursor/mcp.json or a project-level .cursor/mcp.json file with a specific JSON structure. OpenCode follows a similar pattern, requiring edits to ~/.config/opencode/opencode.json with a 'remote' type specification, followed by a CLI authentication command opencode mcp auth svglab. The guide provides a universal test to verify connectivity across these platforms. Users are instructed to prompt the AI to check their SVG Lab account status, create a project named 'Setup test,' and generate a 64x64 paper plane icon. Success is defined by the assistant correctly identifying the account, generating the asset (which may take 10-30 seconds), and returning a project link. For clients that fail OAuth, the guide recommends generating an API key starting with svl_ and passing it as a Bearer header.

Key Takeaways

  • Configuration fragmentation persists: Each major AI client requires a unique, non-transferable setup process for the same MCP server.
  • OpenAI gates MCP access: ChatGPT requires paid subscription tiers and 'Developer mode' activation, unlike the more open CLI tools.
  • File-based configs dominate dev tools: Codex, Cursor, and OpenCode rely on manual JSON or TOML edits, increasing the risk of syntax errors.
  • OAuth is not universal: Some clients require manual API key injection via Bearer headers when OAuth flows fail or are unsupported.

The Bottom Line

MCP standardizes the pipe, but not the plumbing; until clients adopt a unified configuration experience, developers will continue to drown in fragmented setup instructions.