Answers only what it knows
The company's own frequently asked questions, scheduling and booking. Anything else, it says it didn't understand and, the second time, calls a human. It never makes up a price, a diagnosis or a diet.
AI Customer Service area · small businesses
Six sectors, one design: scheduling or booking, automatic confirmation, a waiting list, frequently asked questions, a handover to a human on Telegram, and health and privacy rules built in. Here you will find the system that already runs (clinic), five ready-made kits to implement in your own environment (restaurant, hotel, gym, physiotherapy and nutrition) and the method and tools used to build all of it with agents.
All open and free. Guides in PT, EN and ES. Dependency-free Python, Docker for a VPS, WhatsApp through the Evolution API and the team on Telegram.
01 · What already runs
A complete clinic system, a framework that turns a website into an assistant, and an open CRM for WhatsApp.
Online scheduling, confirmation 48 h ahead, a waiting list, scheduled follow-ups, frequently asked questions, emergencies (192, Brazil's emergency number), a human queue on Telegram, campaigns with a professional-council lock (CFM, CFO, COFFITO: Brazilian professional councils) and LGPD (Brazil's privacy law). It exports the month's numbers to the X-Ray. Checked on a clinic the agent had never seen and in a real docker build.
The company's website and knowledge base become a chat agent on WhatsApp, Telegram, email or a widget: it answers citing its source, schedules, performs services and hands over to a human.
Open the guide → 💬 wacrmAn INEMA guide to an open, self-hostable CRM for tracking and re-engaging customers who chat on WhatsApp.
Open the guide →02 · Kits by sector
Each kit comes with the full specification, 93 to 99 frozen acceptance tests, the guardrails and the prompts for Claude Code or Codex to implement the system in your environment. The system itself is not built yet: it is the same path that took the clinic to 93 of 93.
Tables and rooms, shifts, the smallest table that fits, confirmation, no-shows, a waiting list, large groups with approval and your own customer base outside the marketplace.
Open the guide → 🏨 Hotel / inn · 98 testsPer-night availability without overbooking, seasonal rates, a direct-booking discount, pre-check-in, iCal with Booking and Airbnb, and the source of each booking to measure the commission avoided.
Open the guide → 🏋️ Gym · 99 testsPlans and enrollment, classes with spots and a waiting list, check-in, re-engaging members who disappeared, a workout sheet and an animated explanation of 12 exercises.
Open the guide → 🦵 Physiotherapy · 98 testsTreatment plan, a dropout alert, prescribed exercises with an animated demo, reminders, "done / not done / pain from 0 to 10" and an alert to the physiotherapist. COFFITO lock (Brazil's physiotherapy council).
Open the guide → 🥗 Nutrition · 99 testsIn-person or online consultation, a meal plan with versions, a food diary, water and weight with a neutral reply, and alerts that lead to 192 or to CVV (188), Brazil's emergency and crisis lines. The bot never prescribes.
Open the guide → 🩺 Clinic · implementedThe first to go through the whole cycle: plan, frozen tests, long run, independent verification and publication.
Open the guide →03 · The service flow
The company's own frequently asked questions, scheduling and booking. Anything else, it says it didn't understand and, the second time, calls a human. It never makes up a price, a diagnosis or a diet.
A reminder is part of the duty of care in health; a campaign requires consent; "STOP" blocks everything. The campaign text goes through the sector's professional-council lock before it goes out. The built-in rules are Brazilian (LGPD, CFM, CFO, COFFITO, CFN, CREF).
Receives each conversation with a number, replies on top of the message and uses commands such as /fila, /agenda and /encerrar. An emergency arrives flagged as HIGH.
Automated service doesn't replace the front desk: it takes away what is repetitive and gives back what needs a person.
04 · How they were built
Each kit was written by one agent (Fable) and validated by another that hadn't seen the planning. The validators found 9 to 13 problems per kit, 1 to 3 of which would have blocked execution, and fixed all of them before publication. Implementation runs through a long run, and the runtime is what lets the agent use its tools safely.
A verifiable definition of done, tests frozen by hash, a loop with time and memory caps, a stop on stagnation and independent verification at the end. It is the engine that builds each kit.
Open the guide → 🧭 Agent RuntimeIt is not where the service runs: the systems run on their own, in Python and SQLite. The runtime is the folder you copy into the project so Claude Code and Codex can discover and use their tools (API, MCP, CLI, local bridge) with rules on what they may do on their own.
Open the guide → 📚 Long Runs areaLevels of done criteria, guardrails, context bands and the pilot that closed on the first cycle.
Open the area →05 · In your environment
Each kit has docs/DECISOES-ABERTAS.md with the default proposal already applied. "OK on everything" is enough to start.
bash longrun/<folder>/congelar.sh records the hash of the specification, the tests and the verifier.
The headless loop from execucao-longa (Codex through the subscription) or /goal in Claude Code or Codex. It ends only when every test passes.
docker compose up -d, an HTTPS proxy, the webhook for your Evolution and your Telegram bot. Credentials stay only in your .env.
06 · Why these systems
The clinic's leak no. 1: 5 to 12% in a well-run schedule, close to 25% on average. Confirmation, a waiting list and scheduled follow-ups attack it directly.
The marketplace charges a commission on someone who is already the house's customer. Your own booking and a consented customer base win the channel back.
Commission on repeat guests and empty rooms in the low season. Direct booking, an off-season package and iCal with Booking and Airbnb.
07 · Projects and courses
Implemented, 93/93.
Open the guide → Kit · PT · EN · ESPlan with 93 tests.
Open the guide → Kit · PT · EN · ESPlan with 98 tests.
Open the guide → Kit · PT · EN · ESPlan with 99 tests.
Open the guide → Kit · PT · EN · ESPlan with 98 tests.
Open the guide → Kit · PT · EN · ESPlan with 99 tests.
Open the guide → Project · PT · EN · ESThe site becomes a customer service agent.
Open the guide → ProjectOpen CRM for WhatsApp.
Open the guide → Kit · PT · EN · ESWhere a business loses money, in hard numbers.
Open the guide → Method · PT · EN · ESThe agent working for hours, with frozen tests.
Open the guide → Kit · PT · EN · ESRules, bridges and recipes for the agent to use its tools.
Open the guide → App · PT (in Portuguese)Seven templates that turn into an agent's specification.
Open the app →Four tracks to use the kit and build your own bridge and team.
Open the course → Course · PT · EN · ESSafe agents in a small business: the right problem, limits, prompt injection and cost.
Open the course → Course · PT · EN · ESA clinic process specified with agents: appointment confirmation and claim denials.
Open the course → Course · PT (in Portuguese)What changes in WhatsApp for business: Cloud API, CRM and compliance.
Open the course →08 · Related areas
Where a business loses money, in hard numbers.
Open the area → AreaThe agent working for hours without losing its way.
Open the area → ImmersionYour business with an AI agent up and running.
Open → AreaManaging agents and processes in the company.
Open the area →09 · Get started
Read the open decisions, adapt the example to a business you know and run the /goal. When the agent says "done", run the independent verification.
The template for every kit, with 93 of 93 tests and Docker.
Open the guide →How to let the agent build the system for hours, with proof at the end.
Open the guide →Not sure where to start? Talk to us on Telegram.