Ship it dark.
Flip it live.

Open-source feature flags on the Postgres you already have. Release to 1% or to everyone, and switch it off in a click.

$ npm i featurekit-server
if (flags.get("new-checkout", { key: user.id }, false)) {
  // show the new checkout
}

Raise the rollout, hit the switch, or click a user.

What's a feature flag?

A switch around new code. Your team flips it from a dashboard, not a deploy.

  1. 1

    Ship it switched off

    New code goes live, hidden from everyone.

  2. 2

    Turn it on for a few

    Staff first, then a slice of users, then everyone.

  3. 3

    Switch it off in a click

    Something broke? No rollback, no redeploy.

One switch. Six jobs.

Developer? Flip to see the code behind each one.

Gradual rollout

Widen slowly. Nobody bounces between old and new.

flags.js
fallthroughRollout: {
  variations: [
    { variation: "on",  weight: 20_000 },
    { variation: "off", weight: 80_000 },
  ],
}

Kill switch

Off everywhere before the incident channel fills up.

kill.sh
curl -X POST \
  $FK/v1/admin/projects/acme/environments/production/flags/payments-v2/off \
  -H "Authorization: Bearer $FEATUREKIT_TOKEN"

# every SDK, on the next stream event:
# get("payments-v2")  → false  OFF

Plan entitlements

Pro features for Pro customers, on the app versions that support them.

flags.js
rules: [{
  clauses: [
    { attribute: "plan", op: "in", values: ["pro", "team"] },
    { attribute: "appVersion", op: "semVerGreaterThan", values: ["4.2.0"] },
  ],
  variation: "on",
}]

Beta & staff first

Your team sees it a week before anyone else.

flags.js
targets: [
  { variation: "on", keys: ["u-ana", "u-ben", "qa-bot"] },
]

Remote config

Change copy, limits and prices without shipping code.

pricing.ts
const discount = flags.get("loyalty-discount", ctx, 0);
const banner   = flags.get("promo-banner", ctx, "");

Scheduled launch

Goes live at 09:00. Nobody has to be awake for it.

schedule.sh
curl -X POST $FK/v1/admin/…/flags/bf-banner/schedule \
  -d '{ "patch": { "enabled": true },
       "executeAt": "2026-11-27T09:00:00Z" }'

Self-hosted.
Nothing new to run.

featurekit runs inside a Node app you already have. Postgres is the only dependency.

A change made in the dashboard, the admin API or an agent reaches FeaturekitServer, running in your own Node process, which writes config and audit rows to Postgres and compiles a versioned snapshot. Server SDKs receive the full compiled ruleset by stream or poll and evaluate in memory. Browsers, React Native and edge clients receive only evaluated values for one context, over server-sent events.

  1. 1

    You flip a flag

    From the dashboard, the API, an AI agent or a file in your repo.

  2. 2

    It compiles once

    A bad change never replaces the last good one.

  3. 3

    Every app hears it

    Pushed within seconds. No restart, no redeploy.

A flag check is a function call,
not a network call.

Your servers keep the rules in memory, so asking a flag costs nothing.

featurekitin memory60 ns
a network callbest case1 ms

Slowed down 3,000×. The top bar still finishes before you see it move.

flag not found16ns
flag off50ns
fallthrough, no rules60ns
individual target65ns
rule, 2 clauses177ns
segment match185ns
percentage rollout516ns

Node 22, one core, against the code that ships. Full benchmarks

Six steps, in an order you can say out loud at 3am.

Fixed precedence. First match wins; an error returns your default.

  1. 1not in snapshotFLAG_NOT_FOUND
  2. 2flag offOFF
  3. 3prerequisite unmetPREREQUISITE_FAILED
  4. 4individual targetTARGET_MATCH
  5. 5first matching ruleRULE_MATCH
  6. 6fallthroughFALLTHROUGH

Stale beats wrong.
Both beat empty.

Break something. Your app keeps answering.

Your app: still answering

All fine. Changes stream in as they happen.

  1. live
    Streamed rulesThe normal case.
  2. disk
    Saved snapshotBooting while featurekit is unreachable.
  3. memory
    Last good snapshotUpdates stop or a change won't compile.
  4. code
    Your defaultRequired at every call site, so never undefined.

Ships with alerts and runbooks, including what to do when featurekit_snapshot_age_seconds passes 300.

Browsers get answers.
Never the rules.

Your targeting stays on your server. Each user's device only learns its own result.

Your server holds the rules, such as which plans and countries get a feature. A browser receives only the evaluated value for its own user, for example new-checkout is true.

fk_srv_

Server key

Gets the compiled ruleset and evaluates any context. Rejected by the client routes.

fk_clt_

Client key

Browsers and the edge. Values for one context. Rejected by the server routes.

fk_mob_

Mobile key

React Native and Expo. Same client tier, its own key type.

# what a client key receives + evaluated values for one context - rules, segments and rollout percentages # what a client key can ask for in production (secure mode) + the identity your server signed - anyone else's flags

Every change leaves a trail
you can check.

Approvals, schedules and undo, free. Most tools charge enterprise prices for these.

  1. dry run Preview on real users

    See who'd change before anyone does.

    …/flags/:flag/dry-run
  2. request Get a second pair of eyes

    A teammate approves first.

    …/flags/:flag/requests
  3. schedule Pick the minute

    It applies itself.

    …/flags/:flag/schedule
  4. apply Ship it

    No silent overwrites between teammates.

    expectedVersion: 412
  5. audit Prove it happened

    Hash-chained history. Edits are detectable.

    …/audit/verify
  6. revert Undo to any point

    Back to exactly how it was.

    …/flags/:flag/revert

Roll out by asking.

Works with Claude, Cursor and any MCP client. It can only do what its token allows, and every change is audited.

Claude ChatGPT Gemini Cursor OpenCode
you
Roll out cart v2 to 20% in production.
claude
"cart v2" could mean cart-v2-page or cart-v2-checkout. Which one?
you
The page.
set_rollout cart-v2-page, production, 20%
claude
Done. 20% of production users now see cart-v2-page. PRODUCTION CHANGE
03:12, the pager goes off
you
Kill cart v2 page, checkout is erroring.
turn_flag_off cart-v2-page, production
claude
Done. cart-v2-page is off for everyone.
terminal
claude mcp add featurekit \
  --env FEATUREKIT_URL=http://127.0.0.1:3000 \
  --env FEATUREKIT_TOKEN=fk_pat_... \
  --env FEATUREKIT_MCP_WRITE=1 \
  -- npx -y featurekit-mcp

Feature flags for Node.js,
browsers and mobile.

Two SDKs. Anything else speaks two HTTP endpoints.

npm i featurekit-server
import { FeaturekitSDK } from "featurekit-server/sdk";

const flags = new FeaturekitSDK({ sdkKey: process.env.FEATUREKIT_SDK_KEY, baseUrl });
await flags.ready({ timeoutMs: 5_000 });

if (flags.get("new-checkout", { key: user.id, plan: user.plan }, false)) {
  // synchronous, in memory, no network call
}

Vue and Svelte adapters are in the adapters doc, the wire format in the HTTP spec.

The open-source
LaunchDarkly alternative.

No paid edition. No per-seat bill. Every feature is in the repo.

$0per seat, per user, per check
MITuse it commercially
1dependency: Postgres
870+tests, chaos included
featurekit compared with Unleash and Flagsmith
Feature featurekit Unleash Flagsmith
Self-hosted Yes Yes Yes
Control plane Inside your Node app Separate service Separate service (Python)
Change requests Included Enterprise edition Paid plans
Tamper-evident audit Hash chain you can verify Event log Audit log on paid plans
Framework SDKs One client, 20-line adapters Per-framework packages Per-framework packages
License MIT, no paid edition Apache 2.0 core, commercial enterprise BSD-3 core, commercial enterprise

LaunchDarkly is hosted, billed by seats and usage, and does experimentation. If you need A/B analytics today, it or Unleash is the better fit. If you want governance without a sales call, pick featurekit. Every design decision is written down.

Feature flag FAQ

What is a feature flag? +
A switch in your code that turns a feature on or off, for everyone or for chosen users, without deploying again. It's also called a feature toggle.
What is featurekit? +
Open-source, MIT-licensed feature flags you host yourself. It runs inside your own Node.js app on Postgres, with targeting rules, percentage rollouts, a kill switch, scheduled changes, change requests, an audit log and a dashboard.
Is featurekit a free alternative to LaunchDarkly? +
For flags, rollouts and change control, yes. It's MIT-licensed, self-hosted, and costs what your Postgres costs. It doesn't do experimentation or A/B analytics. If you need those, LaunchDarkly or Unleash is the better fit.
Is there a hosted version? +
No. You run it with Node 22+ and PostgreSQL. To try it locally you only need Docker: development mode starts a throwaway Postgres for you. See the deployment guide.
What happens if featurekit goes down? +
Your app keeps answering. The SDK serves its saved snapshot, then its last good one, then the default you passed to get().
Can someone read my targeting rules from the browser? +
No. Browsers only get evaluated values for one user, never the rules, segments or rollout percentages. In production your server signs each user's identity, so nobody can ask for someone else's flags.
Will a deploy undo a toggle someone made at 3am? +
No. Flag definitions live in a file in your repo, but on/off state and targeting are left alone on deploy.
Is there a React, Vue or Svelte SDK? +
One client works everywhere. React, Vue and Svelte bindings are about twenty lines you copy from the docs, so there's no version matrix to manage.

Flip your first flag.

One install, a dashboard at /admin, and no more redeploys to turn things off.

Want help running it?

We deploy featurekit on your infrastructure, migrate flags from LaunchDarkly, Unleash or Flagsmith, and offer priority support. Bugs go to GitHub issues.