Flagsmith

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me a feature flag service — noting first that this product's own core is open source, so running theirs is a legitimate answer.

Build your own when you want the flag evaluation inside your existing database and nothing else to operate. It is genuinely small.

STACK
- Node 20+ with Fastify
- SQLite through better-sqlite3, WAL mode
- Caddy in front

THE DATA MODEL
- flags: id, key, kind, description, default_value, is_archived — kinds are boolean, string, number, json
- environments: id, name, key
- rules: id, flag_id, environment_id, position, condition_json, value, rollout_percent
- overrides: id, flag_id, environment_id, subject_id, value, reason, expires_at
- evaluations: id, flag_id, environment_id, subject_hash, value, at — sampled, not every call
- changes: id, flag_id, environment_id, before_json, after_json, actor, at — **append-only, and the first thing you will want during an incident**

EVALUATION, WHICH MUST BE BORING
- Rules in order, first match wins, otherwise the default
- Percentage rollout from a hash of the flag key and the subject id — **deterministic, so a person does not flicker between variants on every page load.** Random assignment is the bug that makes a rollout look broken
- Evaluate on the server. Sending your rules to the browser leaks your segments and your unreleased feature names
- **Never let a flag lookup fail a request.** Cache the whole flag set in memory, refresh in the background, and return the last known values if the store is unreachable. A flag service that can take down the application it protects has inverted its own purpose

DELIVERY
- One endpoint returning every flag for an environment, with an ETag. Clients poll it every few seconds; that is cheap and needs no streaming
- Bootstrap the values into the page on first render so nothing flashes

SAFETY, WHICH IS WHY THIS EXISTS
- A kill switch that reaches production in seconds, and this must work when your deployment pipeline does not
- Temporary flags carry an expiry date and a list of expired ones is shown loudly. **A codebase with two hundred dead flags is worse than one with none**, and every team gets there without this
- Changes require a note in production, and every change is in the audit table with who made it

USING IT WELL
- Reference each flag from one place in the code, so removal is one deletion
- Record which flags were evaluated in the last thirty days; the ones that were not are the ones to delete
- A per-flag override for a named account, with an expiry, for support

WHAT MATTERS MOST
Deterministic bucketing and never failing a request. One makes rollouts trustworthy; the other stops the safety mechanism becoming the outage.