A small write-up on what changed when we deleted every gate between an AI agent and a working commerce API.
If you wire Claude Desktop, Cursor, or any MCP-compatible runtime today, you can have BuyWhere returning real merchant data in under a minute:
POST https://api.buywhere.ai/v1/auth/register
→ {"api_key": "bw_live_..."}
GET https://api.buywhere.ai/mcp
→ 3 tools, auto-discovered, free
Three tools: search_catalog, lookup_merchant, normalize_product. ~386M products and ~895K merchants across SG / MY / PH / AU, normalized across marketplaces. First 1,000 requests/day are free; above that, transparent per-tier pricing — no opaque 401s.
The four anti-patterns we deliberately don't ship:
-
Sales-gated onboarding. The agent waits for a human to approve an enterprise contract before the first
curlreturns data. We register programmatically; the key is live in the response. - Unlisted endpoints. REST exists but only MCP is documented (or vice versa, or both with stale examples). We document both surfaces in lockstep.
- Broken tier maps. "Free: 10/day" on a docs page, "ERROR: quota exceeded" on the response. Our free tier is real: 1,000/day, no follow-up email required to verify.
- Opaque 401s. Generic auth failure with no recipe. We return a registration recipe in the 401 body when no key is present.
For Claude Desktop specifically, drop this into claude_desktop_config.json:
{
"mcpServers": {
"buywhere": {
"command": "npx",
"args": ["-y", "@buywhere/mcp"],
"env": { "BUYWHERE_API_KEY": "<register from /v1/auth/register>" }
}
}
}
The full write-up of what changed, the trade-offs, and what we got wrong on the first three iterations is on our blog:
https://buywhere.ai/blog/true-zero-human-self-serve-mcp-2026?utm_source=devto&utm_medium=post&utm_campaign=zero-human-mcp-launch&from=dev-community
If you wire it and hit a real bug, I'd genuinely like to hear about it in the comments. Honest trade-offs are also on the post: 1,000/day is generous but not infinite, and enterprise scale (millions of calls/day, custom catalogs, contract-grade SLAs) still requires talking to us.
Top comments (0)