DeskcommCRM: an open-source CRM where AI agents run your WhatsApp sales

If you sell over WhatsApp — an online store, a clinic, a real-estate agency, a course business — the usual options are paying per seat for Kommo or Intercom, or bolting a chatbot onto some open-source CRM. DeskcommCRM merges the two into one repository: an MIT-licensed, self-hosted CRM with agents that answer customers, qualify leads, move deals and follow up on cooling conversations, all inside WhatsApp. What keeps it from being a chatbot in a costume: the agents reach the CRM over MCP, so the funnel, the agenda and the contacts are things they operate, not text they mime. A WeChat Channels creator featured it on September 12; the repo sits at 4,270 stars as of September 30, 2026.

Core features

• Agents that operate the CRM: each agent carries a per-tenant RAG knowledge base (pgvector), organization memory and executable skills; when something needs a human, it opens a case, keeps talking to the customer while it waits, and the whole handoff is audited.
• Two WhatsApp paths: QR pairing through WAHA (multi-number, with anti-ban throttling, jitter and sending windows) or Meta’s official Cloud API with template approval and sync; media goes through private storage with STOP detection.
• Follow-up without sticky notes: cooled conversations restart on adaptive timing, triggered by funnel stage, and a Radar screen lists the threads at risk of dying unanswered.
• Agents that improve themselves: resolved conversations become new knowledge; an AI-evolution screen shows where the agent is getting better and where it errs, and the agent’s own improvement proposals only ship after a human signs off.
• Vocabulary per niche: pipeline terms are configurable — a lead can read as Cliente, Paciente or Buyer, and a won deal as Paid, Booked or Closed; the same core serves e-commerce (with a Nuvemshop integration), clinics, real estate and infoproducts.
• Multi-tenant and LGPD by design: every tenant-aware table runs behind RLS with an isolation test as a CI gate; audit logs are append-only with 5-year retention, and deletion prefers anonymization.

Typical use cases

• Pre-sales and support for a store: the AI takes the first question, explains the process and opens the deal; people handle only the cases the AI opens. Nuvemshop shops plug in directly.
• Appointment-driven businesses (clinics, services): customers book in chat and the agent puts it on the built-in agenda.
• Small teams that want agents on sales without per-seat SaaS fees: it runs on your VPS and the conversations never leave your server.

Quick start

The main path needs a Docker VPS (4 GB RAM recommended), a domain pointed at it, a free Supabase account and one AI key — OpenRouter, Anthropic or OpenAI:

git clone https://github.com/melgarafael/DeskcommCRM.git
cd DeskcommCRM
bash hostgator-setup-kit/install.sh

The installer generates the secrets, applies the database schema, creates the first admin and brings the stack up with HTTPS; you pair WhatsApp by scanning a QR code during onboarding, and you never hand-install Node or build anything. If you don’t have a server yet, a companion script (comecar.sh) runs on your own machine and tells you which plan to buy, using the numbers from the project’s production runbook.

Summary

DeskcommCRM fits businesses that sell by chat — the README says it started as an e-commerce CRM and the community pushed it into clinics, real estate, agencies and infoproducts. If your sales don’t live on WhatsApp, it has little to offer you. It’s written in TypeScript under the MIT license; 4,270 stars with commits the same day, as of September 30, 2026. Three things to watch: the QR route (WAHA) is an unofficial protocol — throttled and jittered to reduce bans, but the risk is yours, and the official Cloud API is the scaling path. Supabase’s free tier takes no backups, so put backup.sh in your cron. And the UI plus most docs are in Portuguese (the README ships English and Spanish versions; the official site deskcomm.com.br is Portuguese too). For a hosted counterpart, see Revor; for a different take on agents handling customer chats, XianyuAutoAgent runs the same idea on Xianyu.