If you are choosing a tool for feature flags, the market splits in two. Some tools are open source and can run on your own servers. Others are hosted services you pay for by usage, seats or both. Several now do experimentation as well as flags, and most have an official MCP server so an AI assistant can work with them.
This guide compares seven options on the things that decide the choice for most teams: licence, whether you can self-host, how pricing works, whether experimentation is built in, OpenFeature support and MCP support. Everything here was checked against each vendor's own documentation, pricing page or repository on 2 November 2026. Plans and prices change often, so we describe how each vendor charges and link to the pricing page instead of quoting numbers that will go stale.
DeployHQ does not sell feature flags. We deploy the code that your flags sit in, which is why we wrote this from the deployment side.
Quick comparison
| Tool | Licence | Self-host | Hosted | How pricing works | Experimentation | Official MCP server |
|---|---|---|---|---|---|---|
| LaunchDarkly | Proprietary (not confirmed) | No platform self-host; a Relay Proxy runs in your infrastructure | Yes | Per service connection and per client-side monthly active user | Yes | Yes |
| Unleash | AGPL-3.0 | Yes | Yes | Per seat plus API requests (cloud) | Yes, via variants | Yes, alpha per its README |
| Flagsmith | BSD-3-Clause, some MIT | Yes | Yes | Per API request and team members | Segments for A/B tests | Yes |
| GrowthBook | MIT, with enterprise directories under a commercial licence | Yes | Yes | Per seat (cloud Pro) | Yes, core feature | Yes |
| Statsig | Proprietary (not confirmed) | None found | Yes | Usage-based on metered events | Yes | Yes |
| PostHog | MIT outside ee/ |
Yes, but officially unsupported | Yes | Per flag request, no per-seat charge | Yes, bundled | Yes |
| Harness FME (formerly Split) | Not confirmed | Not confirmed | Yes | Monthly tracked keys | Yes | Harness has an official server; FME coverage not confirmed |
Not confirmed
means we could not verify it from a vendor page, so check with the vendor.
Choosing a tool for a team that deploys often? Start a free DeployHQ trial and ship flagged code to production on every push, with the flag off.
LaunchDarkly
LaunchDarkly is the hosted incumbent. It charges per service connection (a server-side SDK in one environment) and per client-side monthly active user, and it has a free Developer plan with limits on both. Experimentation is available, though not on its EU instance.
Its OpenFeature providers for Node, Java, .NET, PHP, Python and Ruby are vendor supported. It has a hosted MCP server, and the docs say that server is not available in its federal or EU environments.
Watch for: there is no self-hosted version of the platform that we could find. The Relay Proxy runs in your own infrastructure, but it proxies LaunchDarkly's service and does not replace it. Its pricing page also showed two different figures for client-side usage when we checked, so work out your own cost from the live page.
Unleash
Unleash is open source under AGPL-3.0 and free to self-host. The open source version has limits on projects, environments and flags per instance, and enterprise features such as role-based access control, SSO and change requests are in paid plans. The hosted service is charged per seat plus API requests and starts with a 14-day trial.
Experimentation works through variants, and Unleash has a guide for A/B testing. SDK coverage is broad: its README lists 15 official SDKs and more than 15 community ones. OpenFeature providers for Node, Kotlin, PHP, Python, Ruby, Rust and Swift are vendor supported.
Watch for: AGPL has obligations if you modify the software and make it available over a network. Ask whoever handles licensing at your company before you build on it. The MCP server's README describes it as alpha and says you run it from source.
Flagsmith
Flagsmith is mostly BSD-3-Clause, with a small number of MIT repositories, and you can self-host the open source build or take the supported Enterprise edition. Hosted pricing is based on API requests per month plus team members, and there is a free plan with a small allowance.
Its README says you can use segments to run A/B and multivariate tests. We could not confirm a built-in statistics engine, so check whether it covers the analysis you need. All of its OpenFeature providers are vendor supported, and it runs a hosted MCP server at mcp.flagsmith.com.
Watch for: governance features such as audit and approval workflows sit in the Enterprise licence, as with most open core projects.
GrowthBook
GrowthBook is open core. The code is MIT licensed apart from directories marked as enterprise, which fall under the GrowthBook Enterprise License. You can self-host it for free, or use the cloud: a free Starter plan for up to three users, and a Pro plan priced per seat.
Experimentation is central: its README describes it as an open source product for feature flags, experimentation and product analytics. It lists 24 SDKs and has an official MCP server.
Watch for: if you only want flags, you are adopting a tool built around experiments. That is fine if you plan to experiment, and extra weight if you do not.
Statsig
Statsig is a hosted platform with usage-based pricing on metered events. It has a free Developer plan, unlimited flag checks on that plan, and experimentation. Its MCP server exposes 18 default tools, and the key you supply decides whether the assistant can write or only read.
Its OpenFeature providers are in the community contrib repositories and are not labelled vendor supported.
Watch for: OpenAI announced in September 2025 that it had agreed to acquire Statsig, subject to regulatory review. We could not confirm the current status, so ask the vendor if long-term ownership matters to you. We also could not confirm the platform licence or a self-hosting option.
PostHog
PostHog puts feature flags inside a wider product analytics platform. Flags are billed per flag request, meaning a single evaluation on the client or the server, with no per-seat charge and the first 1,000,000 requests each month free. Experiments are bundled with the same billing.
The code is MIT licensed outside its ee/ directory. You can self-host it, but PostHog says self-hosted deployments are officially unsupported. It runs a hosted MCP server that covers feature flags.
Watch for: it is a good fit if you want analytics and flags in one place, and a lot of platform if you only want flags.
Harness Feature Management and Experimentation (formerly Split)
Split is no longer sold as a separate product. Harness's documentation says Harness acquired Split in June 2024 and that it is now called Harness Feature Management and Experimentation, or FME. It is hosted, includes experimentation, and is billed by monthly tracked keys, meaning the unique keys evaluated in a month. The documentation warns of a coming change from seat-based to active-user pricing.
Harness has a free plan and an Enterprise plan, but we could not read the free-plan limits or a starting price from its pricing page, so ask the vendor. The Split-branded OpenFeature providers for Node, Web, Go, Java, .NET, Swift and Kotlin are vendor supported.
Watch for: we could not confirm the licence, a self-hosting option, or whether Harness's official MCP server covers FME.
OpenFeature is the part that keeps your choice open
OpenFeature is not a tool you buy. It is an open specification, a CNCF incubating project, that defines a standard way to evaluate flags in code. You write your flag checks against OpenFeature and plug in a provider for whichever backend you use.
Most of the tools above publish OpenFeature providers, and for LaunchDarkly, Unleash, Flagsmith, GrowthBook and PostHog some of them are vendor supported. That makes switching tools a change of provider and not a rewrite of every flag check. Statsig's providers are community maintained, so check their state before you depend on them. Our feature flag guide includes a vendor-neutral example.
How to choose
These are our reading of the facts above, not a ranking:
- You must run it on your own servers: Unleash, Flagsmith or GrowthBook. PostHog can be self-hosted but is not supported that way. LaunchDarkly does not offer the platform for self-hosting.
- You plan to run experiments: GrowthBook, Statsig, PostHog, LaunchDarkly and Harness FME all list experimentation as a core feature. Check the statistics each one provides.
- You already use PostHog for analytics: adding its flags avoids another vendor and another bill.
- You need enterprise governance and support: LaunchDarkly and Harness FME target that, and Unleash, Flagsmith and GrowthBook put governance features in paid tiers.
- You have fewer than a handful of flags: you may not need a tool yet. A boolean in your configuration and a thin wrapper does the job, and OpenFeature lets you move later.
Whatever you pick, write your flag checks against OpenFeature if you can. The same guide covers when to build and when to buy.
Using these tools with AI assistants
LaunchDarkly, Flagsmith, GrowthBook, Statsig and PostHog document an official MCP server, so an assistant can create flags, read their state or change a rollout. Unleash has an MCP server in its own repository, which its README describes as alpha and run from source. For Harness we could not confirm that its official MCP server covers FME.
Treat a flag change as a production change. Start with read-only access, keep production changes with a person, and use a separate key per agent. The AI section of our feature flag guide covers this in more detail. DeployHQ's own hosted MCP server is read-only at the time of writing (2 November 2026).
Where deployment fits
A flag tool decides who sees a feature. Your deployment pipeline decides which code is running. The two work together: deploy the flagged code with the flag off, then release by changing the flag. If you are still building the pipeline side, our complete guide to CI/CD pipelines covers it end to end. To see how flags fit with other release patterns, read about canary releases, trunk-based development and the difference between continuous delivery and continuous deployment.
Shipping flagged code to production? Automatic deployments push every merge to main to your servers with the flag off, so you can release on your own schedule.
Frequently asked questions
What is the best open source feature flag tool?
There is no single best one. Unleash, Flagsmith and GrowthBook are all open source and self-hostable, and they differ in licence (AGPL, BSD and MIT with a commercial enterprise part) and in how central experimentation is. Pick by licence fit and the features you need.
Do I need a paid feature flag service?
Not at first. A configuration value and a small wrapper cover a handful of release toggles. Move to a tool when you need targeting, gradual rollouts, experiments, audit logs or SDKs in several languages.
Is Split still available?
Not as a separate product. Harness acquired Split in June 2024 and it is now Harness Feature Management and Experimentation.
What is OpenFeature and do I need it?
OpenFeature is an open standard for evaluating flags in code. You do not need it, but writing your flag checks against it makes changing vendors a matter of swapping a provider.
Questions or feedback? Email support@deployhq.com or find us on X at @deployhq.