Clawviyo

We are about to live in a world with thousands of useful agents — some inside companies, some run by vendors, some that you and I will spin up for an afternoon's work. The problem is that today every agent ecosystem is a walled garden. Your CRM agent doesn't know your support agent exists. Your design agent can't ask your data agent a question. The web's defining trick — discovery — hasn't happened yet for agents.

Clawviyo started as the connective tissue between agents. As we built, we realized the load-bearing problem was on the merchant side: there was no clean way for a business to be reachable by agents at all. That's the layer we shipped first.

Three things. First, a public trust page per merchant — with all the well-known endpoints an agent crawler expects. Second, an MCP gateway: JSON-RPC 2.0 per company, with built-in tools (request_access, check_status, post_inquiry) plus per-merchant custom tools. Third, a triage pipeline that runs on every inquiry — tiered auth, honeypot and disposable-email filters, per-tool qualification scoring against a merchant-defined schema, reputation updates, and email notification.

The merchant dashboard wraps it all: paste a URL, let Claude scrape your profile, walk through a four-step setup (claim → trust page → install badge → see it work), then sit in the inbox while agents start arriving. Embed snippet is a single <script> tag.

Next.js 16 App Router on Vercel; Supabase for Postgres, Auth, and storage. Two Supabase clients: a service-role client for every API route (RLS-bypassing, service-key-only) and an auth client (anon key + cookies) for dashboard pages.

The MCP route is a thin wrapper that parses the JSON-RPC envelope, extracts agent headers, and delegates to a server module that handles dispatch and protocol negotiation. The runPostInquiry pipeline is the heart of the system: tiered-auth → filters → qualification → reputation → notification, with email failure deliberately non-blocking so the agent always gets a response.

Discovery surfaces are static-ish — well-known endpoints generated per-merchant from shared builders in lib/discovery (llms-txt, skill-file). Onboarding scrape uses Claude Haiku 4.5 with forced tool-use to return a structured {name, description, industry, logo_url}, with an OG-meta-only fallback when the API key is unset.

Merchant onboarding is a deliberate four-step setup surface. Step 1 — claim a slug. Step 2 — verify and describe (Claude Haiku 4.5 scrapes the merchant's website to pre-fill name, description, industry, and logo). Step 3 — write the trust page (with a live preview of the agent-facing card to the right). Step 4 — see it work, where the dashboard polls for the first inbound inquiry from Claude or ChatGPT.

Most merchants get from "paste URL" to "agent can find me" in under four minutes — which is also the marketing headline.

Single-agent demos are a local maximum. The interesting work happens when many specialized agents collaborate — and that requires a piece of infrastructure that nobody currently sells you, because the major model vendors have an incentive to keep you inside their walled garden.

If we get this right, the analogy is closer to DNS or Cloudflare than to a SaaS product — quietly load-bearing, universally useful, and most powerful when it's invisible.

Clawviyo is the most infrastructural of my current projects — narrower than Pangea, more ambitious in scope. I'm building it with Sophia out of MIT AI Studio. The agent products I want to build for Pangea, MyLogia, and the founder community OS all need a layer like this to exist, so we're building the layer.