From API Key to One-Click Install: Building a Claude Desktop Extension for AgentShare

I've been building AgentShare, a price API for AI agents. It works great, but there was always one friction point: users had to manually edit claude_desktop_config.json to add their API key.

For a developer, it's a 2-minute task. But for a user who just wants to ask Claude "what's the best price for a Raspberry Pi 5?", it's a barrier.

So I decided to build a Claude Desktop Extension (.mcpb).

Why an Extension?

The MCP (Model Context Protocol) ecosystem is growing fast. But the "copy this JSON config" approach isn't sustainable for non-technical users. Anthropic's Desktop Extensions solve this — one click to install, secure API key storage, no manual JSON editing.

The Technical Journey

AgentShare's MCP server is remote (Streamable HTTP) , not a local stdio server. This made the extension packaging trickier.

The solution: Use mcp-remote as a bridge inside the bundle:

{
  "server": {
    "type": "node",
    "mcp_config": {
      "command": "node",
      "args": [
        "${__dirname}/server/index.js",
        "https://agentshare.dev/mcp",
        "--header",
        "X-API-Key:${user_config.api_key}"
      ]
    }
  }
}

The extension doesn't hardcode any keys. When users install it, Claude prompts them to enter their AgentShare API key (free tier available). The key stays in Claude's secure storage.

The Result

After building the bundle, I tested it on Windows and macOS:

• ✅ Drag .mcpb into Claude → click install → enter API key → done
• ✅ price_search, best_offer, and best_offer_under_budget tools work natively
• ✅ freshness_status and data_age_seconds give Claude transparency about data staleness

What's Next?

We've submitted the extension to Anthropic's Desktop Extensions Directory. No guarantee of acceptance, but it's a milestone worth sharing.

The .mcpb file is production-ready today. Anyone can install it manually:

1. Download from GitHub
2. Drag into Claude Desktop
3. Get a free API key from agentshare.dev/pricing

The MCP ecosystem is still early. Building good onboarding experiences matters as much as building good tools.

Try AgentShare. Build extensions. Ship things.

#mcp #claude #aiagents `#showdev