You already self-host the chat. This is the rest.

Comparison

Same models, more workspace

Open WebUI and LibreChat are the standard answers for a self-hosted chat interface, and this page will not pretend otherwise — they are good, their communities are far larger than ours, and if a chat UI over your Ollama or vLLM server is the whole requirement, install one of them and be happy. Bee Flow speaks the same OpenAI-compatible endpoints those servers expose. The difference is everything an organisation ends up needing around the chat: sign-in from your identity provider, groups and roles, redaction of personal data, automations, knowledge bases that cite their sources, and the audit trail that says who did what.

Where they win, honestly

Both are genuinely open source under OSI-approved licences, where Bee Flow’s Sustainable Use License is source-available — if OSI purity is a requirement, they meet it and we do not, and the licensing answer below does not soften that. Their communities are larger, their ecosystems of extensions are broader, and a chat UI is a far lighter thing to operate than a workspace: if chat over local models is the entire job, they are the smaller, sharper tool.

What gets bolted on after the pilot works

Then the pilot succeeds, and the questions change: who approved this tool having access to the CRM, can the intern see the salary documents in the knowledge base, what left the network last Tuesday, and can we show the auditor. Those answers are identity, roles, redaction and audit — the parts that are hard to retrofit around a chat UI, and the parts Bee Flow starts with. The difference between a chat UI and a workspace is everything that happens after the pilot works.

The workspace parts

What surrounds the chat here

Identity from your directory, not another user table

OAuth SSO with Google, Microsoft Entra ID and Nextcloud; groups synced from Entra ID and mirrored from Nextcloud in near real time; six built-in roles including a Data Protection Officer and an ISMS auditor; TOTP two-factor; and an access audit log designed for GDPR Article 30 evidence. Sharing follows the directory — a knowledge base shared with a synced group tracks the directory, with no second membership list to maintain.

Redaction before the model

Once the detector is installed, personal data across 21 categories is blocked or tokenised on your own CPU before a prompt reaches any provider, and restored in the reply. No GPU needed, no third-party detection API, and an unreachable detector fails closed rather than quietly open.

Automations, not just conversations

27 step types, six trigger kinds — schedules, webhooks, forms, app events, manual, or an agent calling a routine as a tool — with approval gates, dry runs and per-step history. Built by describing what you want, or by hand on a canvas.

Knowledge with receipts, meetings that write themselves up

Knowledge bases index your documents and answer with citations to the passage used, scoped per group. Meeting notes arrive as diarised transcripts with speakers named, summaries and action items — with the transcription engine self-hostable on your own GPU. Usage monitoring shows which models, which people and what it cost, with a local model reporting zero.

Our boundaries

Said before you ask

Side by side

Chat UI or workspace, in one table

Both projects move quickly — their documentation is authoritative for their column, and this table describes the built-in paths, not what a determined admin can assemble.

Questions

What self-hosters ask us

Does Bee Flow work with Ollama or vLLM?

Yes — point it at any OpenAI-compatible endpoint, which is exactly how those servers present themselves. There is no special Ollama integration and none is needed: the endpoint URL is the integration, and the same goes for llama.cpp, LM Studio or anything else speaking that dialect.

Is Bee Flow open source like Open WebUI?

Source-available, and we would rather be precise than borrow the word. The server and frontend are published under the Sustainable Use License v1.0: read it, modify it, run it internally without paying — what you may not do is resell Bee Flow as your own hosted service. It is not an OSI-approved open-source licence. We are working toward AGPL-3.0-or-later; today the Nextcloud connector is the only part that carries it.

Why is Bee Flow heavier to run?

Because it is a workspace, not a chat window: PostgreSQL with pgvector, S3-compatible object storage, Redis when you scale past one replica, and optional services for PII detection, search and transcription. Compose profiles mean you start with the core and add services as features earn their keep — but if a chat UI is the whole requirement, the honest advice is that a chat UI is less to operate.

Can we keep Open WebUI next to it?

Yes, and teams do — both speak to the same model servers, so nothing conflicts. A sensible migration is running both: keep the chat UI you like while the workspace takes over the parts that need permissions, redaction, automations and audit. Nothing about Bee Flow requires a big-bang switch.

Do you support multiple providers in one place like LibreChat?

Yes — six built-in provider adapters plus any OpenAI-compatible endpoint, selectable per assistant or per conversation, with cost tracked per model. The difference is not the multi-provider chat; it is the permission model, redaction and audit trail underneath it.

Same models. More building around them.

Connect the model server you already run and rebuild one real piece of work — the difference between a chat UI and a workspace shows up in the second week, so start with the part the chat UI does not do.