A feature flag is a switch. An experiment is a measurement of what the switch did. These four all run on your hardware. They do not charge the same way, and only one of them documents a statistics engine in the open-source tree.
GrowthBook leads because the open-source README names a statistics engine, and the Docker image bundles one. The pricing page’s Pro examples include sequential testing, CUPED, sticky bucketing, URL split tests, and product analytics. Flagsmith is next, a BSD-licensed server whose documented experiment is a percentage split. Flipt is one binary and no database, under a Fair Core license on the server. Unleash’s free edition is one project, two environments, and 5,000 flags.
| What you run | Free self-host, in the vendor’s words | License | |
|---|---|---|---|
| GrowthBook | One Docker image, plus MongoDB | Unlimited users | MIT, with some directories under a commercial license |
| Flagsmith | Docker Compose, with PostgreSQL | The README calls the open-source build free | BSD-3-Clause |
| Flipt | One binary, no database by default | Free forever | Fair Core License on the server |
| Unleash | Docker or Node, plus PostgreSQL 15 or newer | 1 project, 2 environments, 5,000 flags, no seat cap | AGPL-3.0 |
Scores
1–5 per criterion · select a score for its reasoning and sources| # | Product | Setup | Free self-hosted edition | Experiments | SDKs | Server license | Overall |
|---|---|---|---|---|---|---|---|
| 1 | GrowthBook | 3 | 5 | 5 | 5 | 4 | 4.4 |
| 2 | Flagsmith | 3 | 4 | 4 | 4 | 5 | 4.0 |
| 3 | Flipt | 5 | 5 | 3 | 4 | 2 | 3.8 |
| 4 | Unleash | 3 | 3 | 4 | 4 | 3 | 3.4 |
How each criterion is scored
Setup
What a documented self-hosted install has to run.
- 5 One binary, and no database by default
- 4 One program plus an optional database
- 3 A container plus one database server
- 2 An application, a database, and a separate cache you must run
- 1 Several services with no single-container path
Free self-hosted edition
What the project says you can run on your own server without a paid plan. Cloud free-tier caps are not treated as limits on the self-hosted build unless the page applies them there.
- 5 Self-hosting is free, and the pages state no cap on users or flags for that build
- 4 The open-source build is described as free, while a separate hosted free plan is capped
- 3 Free self-hosting with a stated cap on projects, environments, or flags
- 2 Free only inside a small hosted quota
- 1 No free self-hosted build
Experiments
How far the documented open-source build goes past an on/off flag.
- 5 A statistics engine for experiments, named in the open-source documentation
- 4 Percentage or multivariate tests the project calls A/B tests
- 3 A percentage split across flag variants, and no experiment statistics engine in the docs we read
- 2 Flags with targeting rules, and no percentage split documented
- 1 On/off flags only
SDKs
Official client libraries, and whether OpenFeature is documented.
- 5 A single published count of 20 or more SDKs, repeated on the pages we checked, or that count plus OpenFeature
- 4 A large SDK list, or OpenFeature plus several languages. Conflicting counts on the project's own pages keep this from a 5
- 3 Several languages, without an OpenFeature provider
- 2 A handful of named languages
- 1 An HTTP API only
Server license
The license on the server you run, not the SDK.
- 5 BSD or MIT for the server, with no commercial directory called out
- 4 A permissive license for most of the code, with a commercial license on some directories
- 3 AGPL or GPL on the server
- 2 A source-available or future-open license on the server
- 1 Proprietary
GrowthBook is the one that treats an experiment as a calculation. The open-source README names CUPED, sequential testing, Bayesian analysis, post-stratification, bandits, and sample-ratio checks, next to feature flags with gradual rollouts. The self-host docs say the Docker image includes a Python stats engine. The pricing page’s Pro examples are sequential testing, CUPED variance reduction, sticky bucketing, URL split tests, and built-in product analytics. The pages do not say which of the README’s methods the free image turns on.
Self-hosting is one Docker image and a MongoDB, or something compatible, which stores logins, cached experiment results, and metadata. The pricing page says the free self-hosted plan has unlimited users, and that cloud Starter is free for up to three users. Flagsmith and Unleash also need PostgreSQL. Flipt’s default install does not.
The homepage says 24+ SDKs. The README says 24, and names React, Python, Android, and iOS. Both numbers are the vendor’s. They agree closely enough to use. The docs say the SDKs evaluate locally and make no network request. SSO and SCIM sit under the heading Open Source, plus, with custom environments and advanced access control. The page does not call that list the free self-host. Cloud Pro is $40 a seat a month, which is $480 a seat a year, on a plan the page describes as up to 30 users. We’d recommend GrowthBook when you want the stats engine bundled in that free image.
What it does
- Feature flags and experiments in one Docker image
- A statistics engine named in the open-source README
- 24 SDKs on the README. The homepage says 24 or more
- Local evaluation, which the docs say means no network request from the SDK
Pros
- Free self-host with unlimited users, on the pricing page
- The Docker image bundles a Python stats engine
- MIT on most of the code
Cons
- MongoDB is required, on top of the container
- Several directories use the GrowthBook Enterprise License
- Cloud Pro is $40 a seat a month, $480 a seat a year
Who should skip:Anyone who will not run a database beside the app.
Flagsmith’s README says most of the platform is BSD-3-Clause, a few repositories are MIT, and the open-source self-host is free, with a separate Enterprise edition.
Docker Compose is the documented path. The compose file sets a PostgreSQL URL. The homepage says a multivariate flag can split a percentage across two or more variations and calls that A/B/n testing. It does not name a statistics engine on that page.
The README’s SDK line is “15+ popular languages”. That is the vendor’s count. OpenFeature is documented as a separate route. The Enterprise image tracks the SaaS build and adds SAML, workflows (change requests and flag scheduling), and LDAP. The only dollar figure on the pricing page is a Start-Up overage of $7 per 100,000 requests, after a grace period. No monthly plan price is stated there, and the pricing page does not say those hosted free-plan caps apply to the open-source build. We’d recommend Flagsmith if the license has to be BSD and a percentage split is enough of an experiment.
What it does
- Docker Compose deployment
- Multivariate flags with a percentage split
- OpenFeature providers
- A separate Enterprise image
Pros
- BSD-3-Clause on most of the platform, MIT on a few repositories
- The README calls the open-source self-host free
- Percentage splits are documented for A/B/n tests
Cons
- PostgreSQL is in the published compose file
- The README's SDK count is 15 or more, not a published list of 20
- SAML, workflows, and LDAP are on the Enterprise image
Who should skip:Anyone who needs those Enterprise controls on the free build.
Flipt v2 is the small install. Those docs say it is a standalone binary, with no database or cache by default and no external services. The percentage rollout is a different document: the v1 introduction defines a distribution as a percentage of entities that get a variant, and the v2 introduction does not repeat that. The homepage says a change in the UI becomes a reviewable Git commit. The v2 docs mark merge proposals as Pro.
The license is the part to read twice. The README says the server is under the Fair Core License, Version 1.0, MIT Future License, and that the Go SDK and generated clients are MIT. The homepage still labels an edition “Open Source” and “Free forever”. The license text is the constraint. Pro is $200 a month, and the README says monthly licenses need a continuous internet connection to validate. Annual is $2,000, which the README says saves $400, and that license can be a file for an air-gapped machine. Merge proposals are marked Pro. We’d recommend Flipt if one binary and Git history are the point, and the Fair Core terms are acceptable.
What it does
- A single binary
- No database or cache by default
- The homepage says a UI change becomes a reviewable Git commit. Merge proposals are Pro
- OpenFeature and OFREP
Pros
- Nothing else to install for a default v2 server
- The homepage says the open-source edition is free forever
- Every change is a Git commit, in the homepage's description
Cons
- The server is Fair Core License 1.0, MIT Future License. The Go SDK and clients are MIT
- Merge proposals are a Pro feature
- Pro is $200 a month, or $2,000 a year
Who should skip:A team that will not run a server under the Fair Core License.
Unleash’s free edition is a number, not a mood. The pricing page says open source is free to self-host, with no limit on seats, and with a limit of 1 project, 2 environments, and 5,000 feature flags per instance. The same page says that edition does not include single sign-on, SCIM, granular role-based access control, or change-request approvals. The README lists A/B testing. That line does not describe a percentage split or a result the test computes.
Pay as you go is $75 a seat a month, $900 a seat a year. A seat is a user of Unleash, not a user of your application. That sentence is on the pricing page, and it is the one that stops the $75 figure from being read as a charge per customer.
Docker or Node.js, plus PostgreSQL. The example uses Postgres 17, or any 15 or newer. By default the instance checks version.unleash.run and sends an anonymous instance id. The docs state that.
SDK counts do not match. The homepage says 25+ official SDKs. The README says 15 official client and server SDKs plus more than 15 community SDKs, and another line says 12 official and 10 community. All three are the vendor’s figures. We’d recommend Unleash when one project and 5,000 flags cover the work and the server license has to be AGPL-3.0. A team that small could also use GrowthBook or Flagsmith. Those pages do not state this cap. We’d recommend GrowthBook when the free self-host has unlimited users and you want the stats engine bundled in the image.
What it does
- Docker or a Node.js process
- PostgreSQL 15 or newer
- A/B testing listed in the README
- A free edition with a project cap
Pros
- No seat cap on the open-source edition
- AGPL-3.0
- The $75 price is per Unleash user, and the page says so
Cons
- Free caps: 1 project, 2 environments, 5,000 flags
- SSO, SCIM, granular RBAC, and change requests are excluded from that edition
- The homepage and the README disagree on how many SDKs exist
Who should skip:A team that needs more than one project before it pays.
How to choose
- You want the free image that bundles a stats engine: GrowthBook. Sequential testing and CUPED are described as Pro.
- The server license should be BSD, and a percentage split is enough: Flagsmith.
- One binary, flags in Git, and Fair Core is acceptable: Flipt.
- One project, 5,000 flags, and an AGPL-3.0 server: Unleash. A team that small could also use GrowthBook or Flagsmith, whose pages do not state this cap.
Our picks
Also considered
- PostHog
This ranking is for products whose documented job is the flag server.
Questions
Is the free plan the same as the free self-hosted build?
GrowthBook’s cloud Starter is three users, and its free self-host is unlimited users. Unleash’s open-source self-host has no seat cap and does have a project cap. Flagsmith’s pricing page describes a hosted free plan. Flipt’s homepage says the open-source edition is free forever, and Pro is a separate license.
Which of these can run an A/B test?
GrowthBook names a statistics engine in the open-source README, and the pricing page describes sequential testing and CUPED as Pro. Flagsmith documents a percentage split across variations and does not name a statistics engine on that page. Flipt’s v1 introduction defines a percentage distribution, and the v2 introduction does not repeat it, so this review does not know whether the current binary has it. Unleash lists A/B testing in the README and does not name a statistics engine.
Does Unleash phone home?
The getting-started docs say a self-hosted instance checks version.unleash.run by default and sends an anonymous instance id, so the UI can mention new releases.
How we scored
Every score cites quotes from each product's own documentation, website and release notes, or from trusted third parties, captured in November 2026. Each criterion counts the same, and a criterion the evidence doesn't cover is shown as "?" and left out of that product's overall score.
SDK counts are the vendors’ own, and Unleash’s homepage and README disagree, so the article quotes both. Flagsmith’s hosted free-plan caps are not applied to the open-source build, because the pricing page does not say they are. Flipt’s “open source” label is the homepage’s. The server license on the README is Fair Core.
