Comparison
Visual building, your infrastructure
Keep the visual canvas — conditions, loops, parallel branches, and an approval gate in front of the irreversible step — and move it onto hardware you own, with a privacy layer the scenario never has to remember to apply. Make is one of the nicest visual builders anyone has shipped, and people who use it enjoy using it, which is rarer in this category than it should be. What it is not is something you can run yourself. This page is about what changes when that starts to matter, and what it costs you.
Try the builder All comparisonsWhere the two differ
The editing experience and the app catalogue. Make’s canvas handles complex branching visually better than most tools manage, the iteration and aggregation model is elegant, and there are far more app connectors than we ship. If you are happy with where your scenarios run, none of this page is a reason to move — switching tools has a cost and enjoying the one you have is a real benefit.
Twenty-four step types: conditions, switches, loops, parallel branches, filters, deduplication, waits, HTTP calls and code. The builder refuses to let you create a cycle, which is the failure you would otherwise find at three in the morning.
Install the detector and personal data is tokenised before any prompt leaves for a model provider, on your own hardware. It applies across chat, assistants and automations at once rather than being wired into each flow.
A scenario that sends an email, issues a refund or deletes a record can pause and wait for a person. Dry runs show exactly what would happen without doing it, and every run keeps a per-step log with sensitive values redacted — so when something goes wrong you can see what it did, not just that it failed.
The 44 built-in integrations are where it starts, not where it stops. Bee Flow speaks the Model Context Protocol, so a scenario can reach anything with an MCP server in front of it — the public ones, the ones your vendors publish, and one you write yourself for an internal system that will never appear in anybody’s app directory. Self-host and you install those; nothing waits on a vendor roadmap.
It is rarely a missing feature. It is a scenario that started out moving rows between two systems and ended up handling names, addresses and case details, at which point someone asks where those fields are being processed and whether a model provider saw them. On a hosted platform the honest answer is a policy document. On infrastructure you run, it is a log you can read.
You take on operations. A container stack, a PostgreSQL to back up, TLS at the front door, and enough disk for whatever knowledge base you build — 2 CPU cores and 4 GB of RAM at minimum, 4 and 8 recommended. That is the same weight as running any other stateful web application, and it is a real cost rather than a footnote. If nobody wants to own it, the hosted tool is the right answer and we would rather you kept it.
What you get for taking that on
Questions
No — there is no importer, and we would rather say that than ship one that half-works. Describing an existing scenario in a sentence and correcting the first draft is usually faster than a conversion tool would be anyway; heavily branched scenarios are the ones that take real time.
Honestly, for pure canvas ergonomics Make has the edge — it is a focused product and it shows. What we add is that you do not have to start on the canvas at all: describe the workflow and it assembles, or let Bee Flow propose automations from work it can see you repeating, then open the canvas to correct.
The self-hosted deployment keeps running. Licence checks are offline against a bundled public key, so an air-gapped install never phones home and no automation stops because a server somewhere was unreachable.
Not as a curated app list, and the number understates what you can actually reach. We ship 44 built-in integrations — Google Workspace, Microsoft 365, Nextcloud, ITSM, payroll, Dutch legal sources — and then the ceiling comes off, because Bee Flow speaks the Model Context Protocol. Any MCP server becomes a set of tools your assistants and automations can call, over stdio or Streamable HTTP: the public ones, the ones your own vendors publish, and the one you write in an afternoon for the line-of-business system nobody has ever built a connector for. When you self-host, installing those is your decision and nobody has to add your system to a catalogue first — the install surface is an enterprise feature, and once a server is installed, using it is an ordinary integration with the same permissions and audit trail as the rest. Where a large catalogue genuinely wins is the long tail of SaaS you want working in ten minutes without writing anything, so check the specific apps you depend on before you switch.
Yes. Point Bee Flow at a model running on your own network — Ollama, vLLM, or any OpenAI-compatible endpoint — and no prompt reaches a cloud provider at all. That configuration is why the privacy layer and the local-model support belong in the same product as the workflow engine.
The one handling data you would rather not explain in an audit. Build it here, then read where every call went.
Try the builder How sovereignty works