Platform
Workflows, internal apps, AI agents, knowledge, meeting notes and a support desk — sharing one permission model, one audit trail and one privacy layer. Build any of it by describing what you want, by dragging it together, or by letting Bee Flow propose it from work it can see you repeating.
Open the app How sovereignty worksStudio
Everything below lives behind one navigation, so a person who learns the chat window can find the rest.
No-code assistants with instructions, tools, knowledge and memory. Share them per group.
Reusable instruction packs an assistant pulls in when the task calls for it.
Hybrid retrieval over your own documents — vector search and keyword search fused, then reranked by a cross-encoder. Local Postgres or a dedicated search service.
Scheduled agent runs and full workflows: conditions, loops, parallel branches, sandboxed code, HTTP calls, approval gates and dry runs. Import an existing n8n workflow and convert it.
Transcription, speaker diarisation, summaries, decisions and action items.
Internal apps with forms, data and permissions — built by describing them.
Full-stack pages with their own database, published from the workspace.
AI-first ticket handling with SLA policies you configure and auto-assignment.
Agents · Skills · Memory · Projects
An assistant is instructions plus context. Bee Flow gives it long-term memory scoped to the user or the project, skills it can pull in on demand, and tools it can call across your connected apps — and any MCP server you point it at. Eleven assistants ship ready to use, including a meeting summariser and a system-prompt designer, so the first useful result does not wait for a prompt-engineering workshop.
See what they can callKnowledge bases
Documents are extracted, chunked with a token budget and de-duplicated before they are ever embedded. At query time, vector search and BM25 keyword search run side by side and are fused, then a cross-encoder reranks what survives. Answers cite numbered sources. If you have not configured an embedding provider, both the embedder and the reranker fall back to models that run in-process on CPU — so a laptop install still retrieves properly.
Live demo
Knowledge bases are the substrate the assistants read from: grouped by category, scoped per group, and counted so you can see what retrieval is actually working with. Seven sample bases here — a handbook, product documentation, a tender answer library, Dutch legal sources.
Sample knowledge bases. No documents are included and nothing is uploaded — the demo has no network access.
Open the real thingRoutines & automations
Workflows trigger on a schedule, an inbound webhook, a manual run, or one of thirteen events in your connected apps — new mail, a changed file, an upcoming meeting. Steps include conditions, loops, parallel branches, filters, aggregation, HTTP requests and code that runs in an isolated sandbox behind an SSRF guard. Approval steps pause the run for a human, dry runs show you what would happen, and every run keeps a step-by-step log with sensitive values redacted.
Try the builderIdentity & access
Everything above shares a single rights model, and it only ever grants: the licence sets a ceiling, the operator chooses what is available, an organisation admin grants from that, and groups grant on top — synced from Microsoft Entra ID, or mirrored live from Nextcloud. Assistants, knowledge bases, skills and apps are shared with the organisation or with specific groups; App Studio apps add row-level access rules inside the app. Every access change lands in an audit log built for GDPR Article 30 evidence, so "who could see this, and since when" is a query, not an archaeology project.
Identity & access, in fullUnder the hood
Nothing in this diagram is a hosted dependency. The optional Python services are separate images you enable per profile.
The security model and the deployment story are the two things most teams check first.
Security & privacy Self-hosting