Your next visitor won't have a mouse
Somewhere in your traffic logs is a visitor that read your pricing page in forty milliseconds, never moved a mouse, and left. It was an AI agent running an errand for a person, and it is not an edge case anymore. This post is about the door we built for it.
Agents already browse — and buy
The two biggest open-source personal agents distribute exactly this behavior. OpenClaw, a foundation-run project with roughly 385,000 GitHub stars, has a registry of more than 5,400 installable skills, including a shopping category — skills that order groceries through a browser and spend with the user's stored credentials. Hermes, from Nous Research at roughly 175,000 stars, pairs an MCP client with hosted browser automation that solves CAPTCHAs and routes through residential proxies. Neither ships purchase guardrails.
The big platforms do the same thing more politely. Amazon's Buy for Me visits merchant sites the merchant never invited it to — the opt-out is emailing Amazon, which merchants discovered with some anger in January 2026. Perplexity's Comet browser automates checkouts and is litigating with Amazon over the right to keep doing it. Whatever you make of each case, the pattern is settled: agents come through the human door, dressed as browsers, and nothing on your side sees them for what they are.
March 2026 settled which side of the counter wins
For a while the endgame looked centralized: checkout would live inside the chat app. OpenAI launched ChatGPT Instant Checkout in September 2025 promising a million-plus merchants — and shut it down in March 2026, with reportedly only a few dozen merchants ever live and Walmart reportedly measuring in-chat conversion about three times worse than clicking through. The industry converged on the opposite shape: discover in AI, buy on the merchant's own site, merchant stays seller of record.
The merchant side won. Which raises the merchant-side question: when an agent arrives at your platform, what exactly does it arrive at?
A door is the smallest honest answer
The agent door is a structured entrance on your own platform. A visiting agent finds a manifest, an llms.txt, and an MCP endpoint that starts with two tools: about, which says what the platform is and what visiting agents may do there, and ask, which answers questions. Where the owner has opened actions, three more appear: list_actions, quote, and confirm. The surface is small on purpose.
- Grounded. Answers come from your platform's own knowledge, with cited sources. When the door doesn't know, it says so — Bubblio never guesses.
- Capped. Asks are rate-limited per door. An agent can be curious; it cannot be a load test.
- Observed. Every visit is a real conversation in your dashboard, transcript and all, marked as agent traffic. Things agents wanted that the door couldn't do land in their own feed: agents tried to…
- Opt-in. The door ships closed. Nothing is exposed until you flip it on, and closing it again doesn't break the URLs.
What a door deliberately cannot do: move money. Actions shipped — each one opt-in per action, run through quote then confirm, executed exactly once, with a duplicate confirm replaying the original receipt instead of running twice. Payment rails are still a later rung, and Bubblio never holds funds.
The Pixalia example
Pixalia, one of the first platforms running Bubblio, sells credit packs — the kind of offering that never fits a product feed, because the useful answer depends on what the buyer is trying to do. Today, a user who tells their agent "find out what topping up my Pixalia credits would cost" gets whatever that agent can scrape and infer. Through a door, the agent asks — pack sizes, prices, whether credits expire — and gets Pixalia's own answer with the source cited, in one round trip, logged in Pixalia's dashboard.
The next rung is built now: an agent with a key can act, not just ask. It quotes the exact call — generate this image, spend these credits — then confirms, and the action executes exactly once against the platform's own backend, with the owner opted in per action and every step receipted. The key isn't ours: the user mints it in their own account on the platform, and the platform's dashboard is the only place an action can be opened — a deploy never opens one. Paying with new money, on the platform's own Stripe, is the rung after this one — not built yet, and we won't pretend otherwise. The order matters: watch what agents want first; let them act second; let them spend last.
Agents are coming through the human door either way. A door built for them is the version where you get to watch.
Open yours
The agent door is live for Bubblio customers now: wire the one signed callback route from /docs/quickstart, flip one toggle in the dashboard, and watch what agents ask. When you're ready for the actions rung, declare them in your code and open them from the dashboard — same guide. The public page at bubblio.dev/agents has the endpoints, the MCP details, and the live directory of open doors — an honest list, currently short. If your platform should be on it, request access.