dif

dif keeps feature flags and experiments where the code lives: one markdown file per flag or A/B test in a dif/ folder, versioned in git like everything else. There is no separate dashboard to keep in sync and no config drift between the flag service and the branch that reads it.

How it works

Install the CLI (npm install -g @dif.sh/cli, published under MIT), run dif init to scaffold the folder, and dif new to draft an experiment. dif build compiles the markdown into a typed TypeScript client you import in your app, so a flag that does not exist is a compile error rather than a silent false. dif conclude closes an experiment and writes the outcome back into the repo.

What the format covers

The same markdown declaration handles plain feature flags, staged rollouts, holdouts and full A/B tests. Audience attributes — country, plan, returning visitor — are declared once and referenced by name, so no customer data is committed alongside the experiment. Each surface (a screen or entry point) keeps a running log of what concluded tests actually taught you, which is the part most flag tools throw away.

Built for coding agents

dif generates a context.json describing every active experiment, so an agent starting a session knows which flags are live, which are mutually exclusive, and which code paths are currently under test. The CLI also does conflict detection: exclusion groups stop two simultaneous tests from hitting the same users and quietly poisoning each other's results.

Pricing

The CLI and SDK are free. Dif Cloud is $50/month — including 1 million events and unlimited seats — and adds live dashboards with exposures, lift and confidence intervals, a shared metrics catalogue, decision history, and drafted conclusions that arrive as pull requests, so git stays the source of truth.