Embeddr

Dating for AI agents. Find your nearest neighbor.

A workshop app for runtime control with Unleash and an AI coding assistant, in Java.

The backend is a small Spring Boot service that evaluates every feature flag server-side through one wrapper, serves Auto-Rizz opening lines behind the auto-rizz flag, and reports Ick taps to Unleash as an impact metric. The frontend (Vite, React, Tailwind CSS v4) holds no flag logic: it renders whatever the backend returns and polls, so turning a flag off stops real server-side work and the running app reflects it within seconds, without a redeploy.

Run it

Prerequisite: a JDK, version 17 or newer. Gradle uses the one you have and downloads its own Node for the frontend into the project folder, so nothing is installed system-wide. You also need an Unleash instance with a Backend token.

cp .env.example .env        # then fill in the instance URL and backend token
./gradlew precheck          # JDK, frontend build, tokens
./gradlew bootRun           # http://localhost:8080

To connect an AI assistant, add your personal access token to .env and run ./gradlew configure, which renders the MCP config for every supported assistant.

On Windows run these in Git Bash, which comes with Git for Windows, and they work as written. In PowerShell or cmd use .\gradlew.bat in place of ./gradlew.

Without a .env the app still boots, every flag evaluates to off, and the app renders the plain Embeddr experience.

After a change to Java code, stop bootRun with Ctrl-C and start it again. After a change under web/, run ./gradlew web in a second terminal; the running app serves the new build, reload the browser.

The flags

Every flag is evaluated in the backend with the session id as context.

Every flag name the app uses is a constant in flags/Flags.java, and every impact metric name in flags/Metrics.java, spelled exactly as in Unleash. Names are unique across a whole instance, so in the workshop each attendee creates their flags with their project id in the name, early-access-melinda-f, and puts that name in Flags.java.

Flag	What it does	Setup
early-access	The waitlist. Off shows the coming soon page to everyone; the invite code a visitor types is the Unleash userId. Wired up by hand in part 2 of the workshop.	Release flag. Target a single invite code to let one person in, then a gradual rollout with userId stickiness to open the doors in slices.
light-mode	Puts a theme toggle in the header. The light or dark choice is the user's own, per session; the flag only decides whether the control exists.	Release flag, gradual rollout at 100%, no targeting.
match-layout-experiment	A/B on how matches are browsed: variant list keeps the card grid, variant deck shows one profile at a time. Match taps feed the match_count_list and match_count_deck impact metrics.	Experiment flag, gradual rollout at 100% with variants list and deck at 50/50, default stickiness.
auto-rizz	Serves an opening line on every match card. Ick taps feed the thumbs_down_count impact metric that a safeguard watches.	Release flag with a release plan from the progressive percentage-based rollout template, walked to the 100% milestone. Safeguard on the metric, optional.

Embeddr Premium is intentionally absent. It gets built live during the workshop, behind a flag the assistant creates, against a payment provider that is down on purpose.

The pieces

Path	What it is
src/main/java/io/getunleash/embeddr/flags/	The flag wrapper. Flags and Metrics hold every flag and metric name. FeatureFlags is the interface controllers use. UnleashFeatureFlags is the only class that imports the Unleash SDK. OfflineFeatureFlags answers when nothing is configured.
src/main/java/io/getunleash/embeddr/api/	One controller per endpoint. Each asks the wrapper a question and returns what the frontend should show.
src/main/java/io/getunleash/embeddr/content/	The opener line bank.
src/main/java/io/getunleash/embeddr/payment/	The payment provider, down on purpose.
web/	The frontend. No Unleash SDK, no flag logic, no token. Built by Gradle into web/dist.
tooling/assistants/	MCP config templates, rendered by ./gradlew configure.
WORKBOOK.md	The attendee workbook. Start here at the workshop.
FALLBACKS.md	Plain commands for every assistant-driven step.
docs/AGENDA.md	Run of show and preparation checklist for presenters.
AGENTS.md, CLAUDE.md	Guidance for AI assistants working in this repo.

Endpoints

Endpoint	Flag	Returns
GET /health		{ ok: true }
POST /access	early-access	{ access }
POST /theme	light-mode	{ themeToggle }
POST /layout	match-layout-experiment	{ layout }, the variant name or null
POST /match	match-layout-experiment	counts the tap under the session's variant
POST /opener	auto-rizz	{ opener }, a line or null
POST /ick		counts a thumbs_down_count tap

Every POST takes { sessionId, userId? } plus endpoint fields and answers 400 without a session id.

Deploy

The Dockerfile builds frontend and backend into one jar; the container serves both from one port. Set UNLEASH_URL and UNLEASH_BACKEND_TOKEN. railway.json wires it up for Railway; PORT comes from the platform.

Behind a corporate proxy

The Gradle wrapper, Node and npm each download something on first run. Point them at your proxy or mirrors:

• Gradle: systemProp.https.proxyHost and systemProp.https.proxyPort in ~/.gradle/gradle.properties.
• Node download: set distBaseUrl in the node block of build.gradle to an internal mirror of nodejs.org/dist.
• npm: registry= in web/.npmrc.