Self-host in one script, or run in the EU cloud
Install Bee Flow free on your own servers with Docker Compose, or let us run it for you in EU data centres.
Install free
Talk to us about hosting
Self-host or managed cloud
The same product either way. What differs is who runs it.
| Self-host | Managed EU cloud | |
|---|---|---|
| Who runs it | You or your IT partner, with Docker Compose. | Bee Flow, in EU data centres. |
| Licence key | Free needs none. | Arranged with us. |
| Updates | Run ./selfhost.sh --upgrade when you choose. | Handled for you. |
| Models on your own network | Yes, through the self-hosted runtimes. | Ask us what fits your setup. |
| Backups | Yours: PostgreSQL and object storage. | Handled for you. |
| Calls home | Nothing by default. Air-gapped installs keep running. | Not applicable: we run it. |
Get started
Install in three commands
The script pulls the public images, generates secrets, waits for the health check and prints your URL and admin password.
What it needs
A Linux or macOS host with Docker and the Docker Compose plugin, plus the parts below. The core profile starts the database and object storage for you.
PostgreSQL with pgvector
Holds your data, including the vectors behind document search.
S3-compatible object storage
Holds files and attachments. The stack ships RustFS by default.
TLS at your own reverse proxy
Bee Flow sits behind the reverse proxy you run, which terminates TLS.
A guard model for personal data
The guard profile loads a model of about 1.2 GB at start and runs on CPU.
Any model, through one interface
Bee Flow has eight provider adapters: Anthropic Claude, OpenAI, Google Gemini, Google Vertex AI, Azure OpenAI, Mistral, Scaleway Generative APIs and EU GPT. Self-hosted runtimes such as Ollama, vLLM, llama.cpp and LM Studio share one adapter. The rest of the product talks to a single interface, so switching models changes nothing else.
Get started
What runs where
Five Compose profiles. The default is core only; add the rest when you need them.
core
core: the workspace
Server, frontend, PostgreSQL with pgvector and object storage. Chat, knowledge bases and automations run here. This is the default.
guard
guard: personal data detection
The guard service finds personal data in messages before they leave. Without this profile a core-only install does no detection.
classify
classify: topic detection
A zero-shot topic classifier that runs on CPU. When the model is set to Auto, it helps pick the model tier for each message, and it powers the 'is about' rule in automation conditions.
local-llm
local-llm: Ollama
Run models on your own hardware with no API keys and no per-token cost. Download a model from the admin settings after the first start.
analytics
analytics: Umami
Adds Umami, a privacy-friendly analytics tool that you host yourself.
Questions IT leads ask
Do we need a licence key?
Not for Free. Starter, Pro and Enterprise use a signed key that your server checks offline. If a key expires, the install falls back to Free and nothing is deleted.
Does it need internet access?
Only to pull the public images and to reach the hosted AI providers you choose. Nothing calls home by default, and an air-gapped install keeps running.
Does it run on Kubernetes?
The images run on Kubernetes, and the documentation has a pattern guide. We do not publish manifests or a Helm chart; Docker Compose is the supported path.
How do updates work?
Set the new image tag and run ./selfhost.sh --upgrade. It pulls the images, recreates the containers, runs the database migrations and checks health. If something fails, it prints the steps to roll back.
Start free, or talk to us first
Install Free in minutes, or tell us where your data has to live and we will say which option fits.
Install free
Talk to us about hosting