MCP for Claude: Remote Image Generation Without Leaving Chat

MCP defines how an AI host and a tool server exchange capabilities (tools, resources), authentication challenges, and tool results. Claude acts as the host. AiWings acts as a remote server at an HTTPS path such as `/api/mcp`. After OAuth, Claude can list tools and invoke them mid-conversation.

Chat attachments and giant base64 blobs do not scale for product mockups or multi-turn edits. A proper connector can open an upload surface, store references server-side, return IDs, and re-use generation_id for “edit the last one.” That keeps context clean and credits accountable.

Concrete tools differ by product. AiWings-style surfaces often include:

• Create reference upload session (interactive widget or signed URL).
• Generate image from prompt + optional references or prior generation id.
• Read preferences (default model/resolution) and credit balance.
• Return structured payloads Claude can summarize while the UI shows media.

Remote MCP for user data almost always requires auth. Claude follows protected-resource / authorization-server metadata so tokens are user-scoped. Users see an authorize screen naming the client (Claude) and the resource (AiWings). Disconnect revokes the story from both sides when done well.

• Prefer Streamable HTTP paths Claude can call without brittle proxies.
• Avoid permanent redirects on the MCP host (www vs apex). Token and metadata discovery often fails when only one host truly serves.
• Resource metadata should match the host the client actually uses.
• List tools only after a valid bearer token so empty tool lists are treated as real failures, not “features.”

Users never need to speak JSON-RPC. They say “mock this bottle in lifestyle light.” Claude chooses tools, AiWings renders, the chat shows the outcome. MCP is the contract that makes that exchange reliable across many services, not just images.

Read Claude Connectors in plain English