Personal agents got a front door: notes on the Personal Agent Protocol

mengyu2022-sudo — Tue, 27 Oct 2026 05:53:58 +0000

I am not affiliated with any of the companies mentioned. This is a summary of public announcements and reporting as of October 27, 2026.

On October 6, 2026, Sierra and Meta announced the Personal Agent Protocol (PAP): an open standard that defines how a personal AI agent identifies itself to a business, states what it is allowed to do, and carries out tasks on a customer's behalf. Founding collaborators include Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart. A v0.1 specification is planned for later in October 2026, with design workshops and a reference implementation to follow.

The problem it answers

Personal agents no longer just chat. Meta's Muse began operating in September 2026; OpenAI's Dots and Anthropic's Claude Duet followed within weeks. These agents browse websites and fill in forms the way a person would, because that is the only interface most businesses offer them.

That friction is already visible. Within two weeks of launch, Amazon blocked Muse, saying the agent did not identify itself as an agent and that unauthorized agent access violated its terms. Other companies reportedly did the same. Businesses currently have no standard way to tell a customer's authorized agent from a scraper.

What PAP proposes

Authentication on OAuth. The agent proves who it acts for, building on infrastructure sites already understand.
Split control. Customers decide what access their agent receives (for example read-only vs. write); businesses decide which actions agents may perform on their systems.
Visibility. Companies can see when an agent is acting for a customer and what it is doing, instead of guessing from traffic patterns.

Two honest caveats from the announcement itself: payments are not in the first version (identity and permissions come first), and neither OpenAI nor Anthropic is participating yet. PAP is a proposal at v0.1 stage, not an adopted industry standard.

A crowded field

PAP is not the only contender. Google's Universal Commerce Protocol (UCP) covers the commerce journey; the Agentic Commerce Protocol (ACP), co-developed by Stripe and OpenAI, routes purchases through chatbot interfaces; Visa's Trusted Agent Protocol and Mastercard's Agent Pay work at the network layer; Google's AP2 works at the protocol layer with cryptographic mandates. Expect overlap and a shakeout.

What builders can do now

The practical preparation is the same work that makes a site agent-ready under any of these standards:

Expose structured, machine-readable interfaces (MCP servers or OpenAPI endpoints) instead of relying on agents scraping pages.
Design for scoped, revocable access — read and write permissions separated, with the customer's agent identifiable in logs.
Decide your agent policy before agents arrive at scale: what may an authorized customer agent do on your site, and what must stay human-only?

Whether PAP wins or not, the direction is set: agents are becoming a channel, and channels need a front door.

References

Sierra Unveils Personal Agent Protocol in Development With Meta and Partners (Unite.AI): https://www.unite.ai/sierra-unveils-personal-agent-protocol-built-with-meta-and-partners/
AI Agents Just Got Their Own Handshake (Sunporch AI): https://sunporchai.com/blog/ai-agents-just-got-their-own-handshake-here-s-why-it-matters-2026-10-09
Meta, Sierra, Walmart, Stripe and Shopify Back a Shared Standard for AI Shopping Agents (T2C Online): https://t2conline.com/restaurant-discounts-are-becoming-another-monthly-subscription/
Revolut Converts Payment Network Into AI-Agent Rail Ahead of Protocol Shakeout (TechTimes): https://www.Techtimes.com/articles/328876/20261013/revolut-converts-payment-network-ai-agent-rail-ahead-protocol-shakeout.htm

Found an MCP server with an unusual design: "Sakupa" — publishes and runs small business websites from your AI chat, with no account at all

mengyu2022-sudo — Tue, 27 Oct 2026 05:53:16 +0000

Sharing this because the design choices are interesting — I'm not the developer, I just came across it while browsing MCP servers and read through the official docs/manual.

Sakupa (npm: @sakupa/mcp) tackles the "AI can generate a website but a non-technical owner can't publish or maintain it" gap. You add one config block (npx -y @sakupa/mcp@latest) to Claude Code/Desktop, Codex, Cursor, VS Code, Gemini CLI, etc., and then publishing, updates, status checks, custom domains, and contact forms all happen in the chat.

Design points that stood out to me:

No account. There is no signup. Control of a site is proven by a local management credential stored under ~/.sakupa/credentials/; the project folder only keeps a reference. Domains are verified via DNS, payments happen only on Stripe-hosted pages.
Pre-publish check. A read-only analyze tool inspects the output folder first and flags files that shouldn't go public; deploy only runs after the user explicitly confirms. Destructive tools (delete, credential rotate) are preview → confirm → execute.
Recovery path. There's a recover tool (device-based for free sites, DNS-based for subscribed ones), and the tool never decides credential revocation on the user's behalf.
18 tools total, including plans/billing that read the actual subscription state instead of guessing, and an apps tool whose first app is an email contact form.

Pricing (from the official page): free preview for ~30 days, then per-website plans from ¥200/month (Water) up to ¥2,000/month (Business). Plans never auto-upgrade.

Caveats from the docs: static sites only — no backend/database apps. It's also a fairly new project (first npm release July 2026), and the server itself is closed-source, distributed via npm; only the plugin/manifest repo is public.

Links: product page https://sakupa.com/product/ · manual https://sakupa.com/manual/ · plugin repo https://github.com/myerwang/sakupa-plugin

Curious what others think about the accountless/local-credential approach for this kind of tool.