Why Bee Flow
Bee Flow is an alternative to ChatGPT Teams and Microsoft Copilot for organisations that want to decide
for themselves where their data goes.
- It runs where you put it. One command on your own server. Use cloud models (Anthropic, OpenAI,
Google, Azure, Mistral), EU-hosted ones (Scaleway, EU GPT), or models on your own hardware (Ollama,
vLLM, llama.cpp, LM Studio and more).
- Privacy is built in. Privacy Shield finds personal data and swaps it for placeholders before a
message reaches a model, and decides per tool what may leave your organisation.
- Compliance you can show. A Compliance Center with continuous checks and the registers auditors ask
for, covering GDPR, the EU AI Act, ISO 27001 and DORA.
- No code needed. Build assistants, automations and small apps visually, on top of your own documents,
mail and calendars.
- It fits your stack. A native Nextcloud integration, a desktop app for Linux, macOS and Windows, and a
native Android app.
What you can do
Assistants that know your organisation
- Build an assistant for a team or a task: instructions, knowledge, skills, tools and model, on one screen.
- Knowledge bases from files (PDF, Word, Excel, CSV, Markdown), web pages, tables and meeting notes, with
OCR for scanned documents.
- Hybrid search (pgvector plus full text) reranked by a self-hosted cross-encoder. Answers cite their
sources.
- Notebooks: a document you write together with the AI, next to the sources it draws from.
- Connect Google Workspace, Microsoft 365, Nextcloud, GitHub, web search and any server that speaks the
Model Context Protocol (MCP).
Automations you can watch
Automations chain triggers, AI steps, logic, people and your apps into one flow, and show every run step by
step. Steps that act on your behalf can wait for a person to approve them. Start an automation on a schedule,
from an event, a form or a chat.
Meeting notes with the decisions in them
Record or upload a meeting and get a transcript with speakers, a summary, the decisions and the action
items, each linked to the moment it was said. Transcription runs on your own server (WhisperX) or through
a cloud API.
Compliance and oversight
And more
- Privacy Shield (shown at the top): detection of names, addresses, bank and ID numbers, credentials
and more with a model you host yourself, placeholders instead of real values, and a check with the
person before data leaves for an outside tool.
- Support inbox (beta): a shared inbox where the AI drafts replies and flags what needs a person.
- Studio: forms, data tables, documents (quotes, invoices, letters as PDF or Word), web pages and small
apps, all connected to your automations.
Quick start
Run your own Bee Flow server with one command. It pulls the public images from ghcr.io/bee-flow (no
login needed), generates secrets, starts the core stack and prints your URL and admin password.
git clone https://github.com/Bee-Flow/Bee-Flow.git
cd Bee-Flow
./selfhost.sh
Free to start. The Community tier needs no licence key: chat, knowledge bases, multiple users and the
Nextcloud connector. Premium features such as automations, meeting notes and guardrails unlock with a
licence under Admin → Licence. See beeflow.nl.
Using Nextcloud? Point the Bee Flow connector at your server from Nextcloud → Settings →
Administration → AI → Bee Flow, choose Self-hosted server and enter your URL.
The full guide is in the self-hosting documentation.
The server image runs as node (uid 1000). Two consequences:
- It still drives Docker over the mounted socket (installing the PII guard, the browser runner and the
security scan), so it must be in the socket's group. The published image is built with
DOCKER_GID=999, the usual gid on Debian and Ubuntu. If stat -c %g /var/run/docker.sock says
otherwise, add group_add: ["<that gid>"] to the server service, or rebuild with
--build-arg DOCKER_GID=<that gid>.
- Upgrading an install that ran the old root image: its
beeflow-data volume is root-owned, and so is
whatever the old image wrote into components/. Hand both to the new user once before starting the new
image: docker compose run --rm --user root server chown -R node:node /app/data /components. The same
is needed whenever components/ is not owned by uid 1000: a first docker compose up without
./selfhost.sh, ./selfhost.sh run with sudo or by another user, or rootless Docker. Until then the
server cannot install the npm dependencies of the shipped components there at boot, and those
components fail to run.
| File | Purpose |
|---|
docker-compose.from-registry.yml | Self-host from the prebuilt public images. This is what ./selfhost.sh uses. Profile-based (core, guard, classify, local-llm, analytics). |
docker-compose.yml | Builds server, agent-hub, guard-service and the browser sidecar from this checkout. |
docker-compose.dev.yml | Development with the source mounted into the containers for hot reload. |
docker-compose.hub.local.yml | Bee Flow-internal, local-only override on top of the registry stack for module-hub development — flips the server to NODE_ENV=development; never use on a real deployment. |
docker-compose.security.override.yml | Opt-in layer that enables the security-scan worker on the registry stack. |
Each file starts with a header comment that explains its use.
Develop locally
For day-to-day work you only need the two core services, the React frontend and the Node.js backend, plus
a PostgreSQL database with pgvector.
Prerequisites: Node.js 22 (>= 22.12, see .nvmrc) with npm 10+, Docker for the database container,
Git, and Python 3.10+ on Linux and macOS for the install script.
git clone https://github.com/Bee-Flow/Bee-Flow.git
cd Bee-Flow
python3 install-local.py # Windows: powershell -ExecutionPolicy Bypass -File .\install-local.ps1
npm run dev:all # backend on :3101, frontend on :5276
The installer installs the npm dependencies, starts a pgvector container named beeflow-postgres on
host port 55432, writes a .env with a fresh SESSION_SECRET and runs the database migrations.
- Run the two services in separate terminals with
npm run dev:server and npm run dev:frontend.
- Start and stop the database with
docker start beeflow-postgres and docker stop beeflow-postgres.
- Without Docker, install PostgreSQL 15+ with
pgvector yourself and create a database beeflow_tasks
owned by user beeflow (password beeflow). The script notices that Docker is missing and skips that
step.
- The ports (
55432, 3101, 5276) are deliberately unusual so they don't clash with other servers you
may run. Change them in .env if you like.
- Optional services (object storage, Redis) are not needed for the core experience. Start one with
docker compose -f docker-compose.dev.yml up -d <service>.
See CONTRIBUTING.md for the test commands per module and how pull requests are checked.
Models and providers
| Where the model runs | Providers |
|---|
| Cloud | Anthropic Claude, OpenAI, Google Gemini and Vertex AI, Azure OpenAI, Mistral |
| EU-hosted | Mistral (EU endpoint), Scaleway Generative APIs, EU GPT |
| Your own hardware | Ollama, vLLM, llama.cpp, LM Studio, SGLang and other OpenAI-compatible runtimes |
Embeddings, reranking, PII detection and transcription can all run on your own server, on CPU or GPU.
Clients
| Client | Runs on | What it is |
|---|
Web (agent-hub/) | any browser | The workspace, served by your own server. |
Desktop (desktop/) | Linux, macOS, Windows | A native shell around the same web app: tray icon, a global quick-ask shortcut, notifications, beeflow:// links, and a bridge to the Nextcloud desktop app so attached files are referenced instead of copied. See desktop/README.md. |
Android (mobile/) | Android phones | A native app with its own phone-shaped design, not a web wrapper. See mobile/README.md. |
Architecture
- Frontend: a React single-page app (Vite, Tailwind CSS).
- Backend: Node.js and Express for the API, streaming chat and automations.
- Storage: PostgreSQL with
pgvector, S3-compatible object storage (RustFS) and Redis.
- Sidecars: small Python services for PII detection and topic classification. Each is optional.
The repository layout is described in REPO-STRUCTURE.md.
Security and privacy
- Personal data stays inside. Privacy Shield replaces personal data with placeholders before a
prompt reaches a model and controls per tool what may leave. Detection runs on your own server.
- Encryption at rest, per organisation (Enterprise licence): choose managed keys, which an
administrator can recover, or zero-knowledge, where only the user can decrypt their content. It is off
by default. Built on AES-256-GCM, Argon2id and OPAQUE.
- Sign-in: passwords, Google, Microsoft and Nextcloud sign-in, SAML (Enterprise), and two-factor
authentication with an authenticator app or a security key.
- Audit trail: sign-in and access logs, and a hash-chained evidence log in the Compliance Center.
Found a vulnerability? Please report it privately, as described in SECURITY.md.
Licence
Bee Flow is fair-code. You may use it, change it and run it for your own organisation. Offering Bee
Flow itself as a paid hosted service to others needs a separate agreement. The server, frontend and
clients use the Sustainable Use License; the Nextcloud connector uses AGPL-3.0-or-later. See
LICENSE.md for the overview per component.
