Self-hosted AI customer support, with the data in your own database
Self-hosted AI customer support that answers customers and acts on orders, on servers you control. The code is source-available and runs with Docker, PostgreSQL and Redis.
Is Hay.chat open source?
Not quite, and the difference is two rules. Hay.chat is source-available under the Hay Community License v1.0: free to use, modify and deploy, including self-hosting on your own infrastructure. You may not sell the software itself or remove the "Powered by Hay" attribution from the chat widget. It is not an OSI-approved open source licence.
If you searched for self-hosted open source AI customer support software, this is probably what you meant: the code is public and free to run for your own customers.
Read the licence
Read the licence
What do I need to run Hay.chat on my own server?
What you need
- Node.js 18 or later, and npm 9 or later
- Docker, for PostgreSQL 16 (with pgvector) and Redis 7
- An OpenAI API key, which embeddings always use
- Two auth secrets of at least 32 characters
The dashboard opens on localhost:3000 and the API on localhost:3001. The last command starts a development server, so this gets you an instance to evaluate. Running it in production is yours to design. The technical docs cover the architecture, plugins and API.
From the hay-core README, as of 21 October 2026
# Clone the repository
git clone https://github.com/hay-chat/hay-core.git
cd hay-core
# Start infrastructure
docker compose up -d
# Install dependencies
npm install
# Configure environment
cp .env.example .env
# Edit .env with your settings
# Run database migrations
cd server && npm run migration:run && cd ..
# Start development
npm run devWhat stays on my servers, and what goes to an LLM provider?
Your data is stored in your own PostgreSQL. Two things leave the server: the text the AI needs to write each reply, and the text used for search embeddings. There's no fully offline mode today.
On your servers
PostgreSQL 16 with pgvector holds the conversations, customer records, your knowledge base and its search vectors, your playbooks and your settings. Redis 7 handles caching and the real-time features. The app runs wherever you deploy it.
Sent to your LLM provider
To write a reply, the customer's message and the context retrieved for it go to the model: OpenAI on your key by default, any OpenAI-compatible host, or Anthropic, Google Gemini, Mistral or xAI Grok per workspace. Embeddings always go to OpenAI.
Sent to the tools you connect
When the agent looks up an order or creates a refund, the plugin calls that tool's own API (WooCommerce, Stripe, Zendesk and so on) with the credentials you gave it. Connect nothing and nothing goes there.
So self-hosting changes where your data is stored, not whether the model sees the conversation. Want your team to read replies before customers do? Approval is in the same code, and the human approval page shows how it works.
Should I self-host Hay.chat or use the cloud?
Self-host if a contract, a regulator or your own policy says conversations have to live in a database you control, or if you want to change the code. Not sure? Try the cloud first. A month of real conversations tells you more than a weekend with Docker.
Swipe the table sideways to see every column.
| Question | Hay.chat cloud | Self-hosted Hay.chat |
|---|---|---|
| Who runs it | We do. Sign up and you're in the dashboard | You do, on servers you choose |
| Where conversations are stored | In the EU, on every plan. Kept until you delete them or set a retention period | In your own PostgreSQL, wherever it runs. Retention is up to you |
| Where AI replies are generated | By OpenAI in the US, under EU Standard Contractual Clauses | By the provider you configure. OpenAI on your key by default, or any OpenAI-compatible host |
| Embeddings | OpenAI | OpenAI, on your key |
| Updates | Shipped by us. Nothing to install | You pull new code from the repo and run the migrations |
| Support | Email support on Solo, priority support from Starter | Community project: GitHub issues, and we answer emails where we can |
| Cost | Flat monthly plans from €50. 30-day free trial, no credit card | The software is free: no fee, licence key or billing check. Servers and LLM usage are yours to pay for |
| What you can't do | Nothing extra beyond the terms of service | Resell it, run it as a competing hosted service, or remove the "Powered by Hay" mark from the widget |
As of 21 October 2026. Cloud facts are from our pricing and security pages. Self-hosted facts are from the hay-core README and LICENSE.
How does Hay.chat compare with Chatwoot, Zammad and Rasa for self-hosting?
They solve different problems. Chatwoot is a shared inbox with an AI add-on, Zammad is a ticketing system, and Rasa is a framework for building your own assistant. Hay.chat is the AI agent itself, and it can run on top of Chatwoot. For the helpdesks on their own, see our open source helpdesk comparison.
Swipe the table sideways to see every column.
| Tool | Licence | AI out of the box, self-hosted | E-commerce actions | How you host it | Paid self-hosted support |
|---|---|---|---|---|---|
| Hay.chat | Hay Community License v1.0. Source-available, not OSI-approved | Built in. Answers from your knowledge base, follows your playbooks, escalates to your team with the context | Plugins in the repo for Shopify, WooCommerce, Magento and Stripe. Order lookups and refunds, within your playbook rules | Docker for PostgreSQL and Redis, plus the Node.js app | Community project: GitHub issues, and we answer emails where we can |
| Chatwoot | MIT. Community Edition free forever | Captain, its AI agent, answers common questions and helps agents reply. Not included in Community Edition | Shopify integration shows orders and customer info to your agents. For the AI to act, Captain calls HTTPS endpoints you build and host (custom tools) | Docker Compose, Kubernetes, one-click Heroku and DigitalOcean | Premium Support $19 or Enterprise $99 per agent a month, both with Captain |
| Zammad | GNU AGPLv3 | No customer-facing AI agent stated. AI agents sort, route, prioritise and tag tickets. AI summaries and a writing assistant help your team. Can use Ollama on your own hardware | Not stated | Docker, packages for Ubuntu, Debian and CentOS, or from source | Support contracts from €2,999 a year |
| Rasa | Rasa Open Source: Apache 2.0, now in maintenance mode. Current product has a free Developer Edition (one bot, up to 1,000 external conversations a month) | A framework. You design and build the assistant yourself, in code or the visual builder | Not stated. You build any actions yourself | Self-managed on-prem or private cloud, Kubernetes via Helm | Enterprise plan, price on request |
As of 21 October 2026. Every competitor cell is from the vendor's own website, docs or GitHub repo, with prices as shown there. "Not stated" means we couldn't find it on the vendor's pages. Hay.chat's row is from the hay-core README and the plugins in the repo.
Run it yourself, or let us run it
Clone hay-core and self-host it for free, or start on the Hay.chat cloud with a 30-day free trial and no credit card.
- 30-day free trial
- No credit card required
- Set up in under 10 minutes
- Hosted in Europe
Frequently asked questions
Written by Damien Mulhall. Last reviewed against the hay-core README and LICENSE.