What it costs at three sizes

Retool figures are the Business plan, which is the first tier with role-based permissions and audit logs. Magic figures are the Professional cloudlet at $100 a month, or Enterprise at $300 where the extra machine helps. Retool prices verified 19 September 2026.

One honest caveat on the third row. Retool lists external users as free to 50, then $8, $6 and $4 per month at rising volumes, without stating whether the rate is graduated or applied flat at your volume. Read both ways, 500 external users land between roughly $3,000 and $3,100 a month, before the builder seat. The range is theirs, not ours.

Self-hosting is a sales call, not a checkbox

Retool’s pricing page has a Cloud/Self-host toggle. Switch it to Self-host and three of the four plans change to “Plan not available on Self-host.” What remains is Enterprise, priced on request. Their documentation is consistent with this: the self-hosted deployment pages are labelled Enterprise, production deployment is described as Kubernetes via Helm, and Docker is offered explicitly for testing purposes.

That is a legitimate commercial choice, and for a large regulated buyer with a procurement function it may not even register as friction. But it means self-hosting is not something you can evaluate on a Tuesday afternoon. You cannot price it, you cannot start it without a contract, and the smallest unit of commitment is an enterprise agreement.

Magic inverts this. Self-hosting is the primary deployment model and it costs nothing, because the runtime is MIT-licensed. One command brings up the whole stack — database layer, generated APIs, auth and RBAC, file handling, task scheduling and the MCP server — and the hosted cloudlets run exactly the same code, because there is nothing gated to gate.

The open-source claim

Retool maintains public repositories, and several carry the MIT licence. It is worth being precise about what they contain: deployment scripts for running Retool on your own infrastructure, Helm charts for Kubernetes, Terraform modules for cloud provisioning, an RPC library and a React embedding component. Useful engineering, openly published.

None of it is the product. The on-premise repository carries no licence file of its own, and its documented flow is to run an install script that writes a license key into an environment file and then pull a versioned, closed container image from Retool’s registry. The open part is the thing that installs Retool. Retool itself is closed.

For comparison: Magic is MIT in full — the runtime, the plugins, the slots, the dashboard, the CRUD and SQL generators, the auth and RBAC layer, the AI features. Fork it, run it, sell it, never speak to us again. There is no key, no threshold, and no plan that unlocks a feature the self-hosted build is missing.

Magic Cloud vs Retool, honestly

Thirteen rows: nine to Magic, four to Retool, each marked. The visual builder row is the one most people should weigh hardest.

When you should stay on Retool

You need non-developers building interfaces. This is the big one. Retool’s canvas lets an analyst or an ops lead assemble a working, decent-looking tool — tables, forms and charts wired straight to queries — without writing a frontend. Magic’s Web Designer is a genuine WYSIWYG editor, but it edits pages: there is no component bound to a query, so an analyst building an internal tool is still better served by Retool, and it is not close.

You need native mobile. Offline mode, push notifications, biometric authentication and a white-labelled app in the stores are real product surface, and we do not have them.

Your integration list is long and exotic. Retool’s connector catalogue is far broader. If your tool has to speak to fifteen SaaS products on day one, count the connectors before you count the dollars.

Your seat count is small and staying small. Three developers on Team is $30 a month. The per-seat model is genuinely cheap at the bottom; it is growth that makes it expensive. If you are not growing, the arithmetic above never fires.