Counter

Phones are for people. Counter is for agents.

Every store gets one URL that personal AI agents (Muse, Dots, Pickle, Claude…) can talk to. Counter answers from what the store has told it, holds a slot and asks staff to approve anything irreversible, and when it doesn't know something it asks staff once — then remembers.

Live

https://counter-agent.fly.dev/s/drd-yeongdeungpo — Counter for a real Seoul dermatology clinic (Dr. Designer Clinic Yeongdeungpo, built from its public website; demo only, not affiliated). Give that URL to any AI agent: it reads the clinic's facts, asks what isn't covered, and books a slot that staff approve. People who open it in a browser see exactly what agents see.

Run

node app/server.mjs --reset     # Node 18+, no dependencies. PORT=3000 by default
ADMIN_TOKEN=... node app/server.mjs   # when exposed publicly (e.g. a tunnel): staff open /admin?key=<token> once

URL	Who	What
/s/dojagi	AI agents (markdown) / people (web page + chat)	The store's agent contact point
/s/dojagi.md, /llms.txt	AI agents	Facts, open slots, learned answers, API
/.well-known/agent-card.json	AI agents	Discovery card
/admin	Store staff (phone) — needs ADMIN_TOKEN when set	Approve bookings · answer unknown questions · memory · settings

Agent API (/api/s/dojagi): POST /ask · POST /book — add "wait":60 to get the staff outcome in the same call · GET /tasks/:id?wait=60 (long-poll). No human UI: people who open the store URL see the same markdown agents read. Browser-only agents use the same actions as plain links: GET /ask?q=… · GET /book?service=…&date=…&time=…&name=…. Task states: input-required (waiting for staff) → completed / rejected.

Demo scenario

1. node app/server.mjs --reset, open /admin on a phone (same Wi-Fi: http://<mac-ip>:3000/admin, or cloudflared tunnel --url http://localhost:3000).
2. Run Claude Code as the tourist's agent:

   claude -p --allowedTools "Bash(curl:*)" -- "You are Muse, a personal AI agent for Emma Clarke, a British tourist in Seoul who uses a wheelchair. Emma wants Rejuran Healer tomorrow (2026-11-10) around 3pm at Dojagi Clinic, and would prefer a female doctor. The clinic's contact point for agents is http://localhost:3000/s/dojagi — use curl. Check accessibility, ask about a female doctor if not listed, then book. Report back to Emma in 3 short lines."

3. Accessibility is answered instantly from the store's facts.
4. "Female doctor?" is unknown → appears in /admin → staff answers, keeps Remember on.
5. Booking request → slot held → staff taps Approve → the agent gets the confirmation number.
6. Ask again from another agent (e.g. just "Do you have a female doctor?") → answered instantly from memory, no staff needed.

Measured on 2026-11-09 with Claude Code: whole flow 24 s end to end (staff responding 5 s after each request). Use cases and scenarios: docs/use-cases.md.

Team

Folder	What
app/ (server.mjs, public/, data/)	Product (build) — all product code lives here
video/	Demo video work — one folder per teammate (Korean)
docs/use-cases.md	Use cases UC-1~5 + demo scenarios — the shared reference for build, demo clips and video (Korean)
docs/measurement.md	Phone vs Counter measurement plan (Korean)
docs/	Meeting notes (Korean)

Before any work: git pull --rebase origin main — see AGENTS.md.