Star
Start here

Why featurekit

Feature flags with a control plane you run. Ship code switched off, turn it on for 1% or everyone, and switch it off in a click. On the Postgres you already have, MIT, with no per-seat bill.

if (flags.get("new-checkout", { key: user.id, plan: user.plan }, false)) {
        // the new checkout
      }

That check is synchronous and in memory: about 60 nanoseconds, with no network call. Behind it, rules are pushed to every server within seconds, rollouts are sticky per user, a kill switch turns a flag off everywhere, and the SDK keeps answering when the control plane is down.

What you get

 What it means for you
TargetingTurn a feature on for staff, a plan, a country, an app version or a list of accounts.
Percentage rolloutsStart at 5%, widen to 100%. Nobody flips between old and new as you widen. How
Kill switchOff everywhere, immediately, with no version check to slow you down.
Change controlScheduled changes, change requests with approval, dry runs, revert, and a hash-chained audit log. How
Remote configFlags can be boolean, string, number or json, so copy, limits and prices change without a deploy.
Flags in gitDefinitions live in a file in your repo. On/off stays with whoever is on call. How
A dashboardAt /admin, served by the same process.
An AI agent"Roll out cart-v2 to 20%" from Claude Code or Cursor. MCP server

Why it's built this way

Checks cost nothing. The server SDK holds the compiled rules and evaluates in memory. A plain check is about 60ns, a percentage rollout about 0.5µs. You can check a flag in a hot loop.

Browsers never see your rules. Client keys get evaluated values for one user, never the rules, segments or rollout percentages. A rule like plan == "enterprise" doesn't leak your pricing plans to anyone with devtools. In production, the user's identity is signed by your server.

It survives outages. When the control plane is down, the SDK serves its last snapshot, then the one it saved to disk, then the default you wrote at the call site. A flag check never throws and never returns undefined.

There's no flag service to run. The control plane is a class you start inside a Node process. Postgres is the only dependency. No Redis, no queue, no vendor in the request path.

Every feature is in the repo. Change requests, the audit chain and secure mode aren't held back for an enterprise edition. There is no paid edition.

What it doesn't do

  • Experimentation and A/B analytics. featurekit decides who sees what; it doesn't measure the outcome.
  • A hosted version. You run it.
  • SDKs in a dozen languages. There's one server SDK (Node) and one client SDK (browsers, React Native, Deno, Bun, edge). Anything else speaks two HTTP endpoints.

If you need those today, LaunchDarkly or Unleash is the better fit.

How it compares

 featurekitUnleashFlagsmith
Self-hostedYesYesYes
StoragePostgresPostgresPostgres
Control planeA class in your Node processSeparate serviceSeparate service (Python)
Change requestsIncludedEnterprise editionPaid plans
Tamper-evident auditHash chain you can verifyEvent logAudit log on paid plans
Framework SDKsOne client, 20-line adaptersPer-framework packagesPer-framework packages
LicenseMIT, no paid editionApache 2.0 core, commercial enterpriseBSD-3 core, commercial enterprise

LaunchDarkly is hosted SaaS and the most complete option, with experimentation and many SDKs, billed by seats and usage. featurekit costs whatever your Postgres costs.

Next