The Free Coding Agent
Built For Work
Built on Solana. Free access to Opus 5.5, DeepSeek v4.1, and other powerful coding models for $MILK holders.
Free Powerful Models
Get a daily allowance to spend on a range of models. Find and test the best coding model for your project completely free.
- OpusΒ 5.5
- DeepSeek v4.1 Flash
- GLM 5.3 Flash
- GPT-6 Luna
Powerful Tools to Build, Preview, and Publish Anything
Build anything from websites, backends, and even Swift apps with FreeMilk's built-in execution environments.
FreeMilk builds Dusk Raid, a 3D first-person shooter made with Three.js, and runs it in the live preview.
FreeMilk builds Calorie Tracker, a calorie tracking iPhone app, and runs it in the iPhone 17 Pro simulator, open at its Progress screen.
FreeMilk builds Honeybun CafΓ©, a landing page for an award-winning cafΓ©, and previews it live.
FreeMilk queries PostHog and writes a report on why day 7 retention dropped after a release, with the HogQL to reproduce it.
100%
Make a 3D first-person shooter in a ruined city at dusk
Which kind of shooter should it be?
This decides the camera, the controls and how big the map is.
First-person, one ruined city block, with drones as enemies- Thought for 4s
- Created index.html +46
- Created game.js +486
- Created world.js +266
- Ran the game with no console errors
I've built Dusk Raid, a first-person shooter in Three.js. It runs right in the preview, so click into it to play.
What's in the game:
- City: low-poly towers at dusk, with rubble you can kick around.
- Controls: WASD to move, the mouse to aim, Shift to sprint and G for grenades.
- HUD: kills, score, vitals and a 30-round magazine.
- Enemies: drones that patrol the streets and chase you once they spot you.
Files are in index.html, game.js and world.js.
Build an amazing iOS calorie tracker that scans meals from photos
Where should this run β a web app in the browser, or a real iPhone app?
A web app I can build and test end-to-end here. A native iPhone app runs in the simulator and feels the most like the real thing.
Native iPhone app in SwiftUI, running in the simulator- Thought for 6s
- Created ProgressView.swift +486
- Created FoodScanner.swift +312
- Created FoodDatabase.swift +1,940
- Launched on the iPhone 17 Pro simulator
Calorie Tracker is built and running. It launched on the iPhone 17 Pro simulator with no crash.
What's in it:
- Today: a calorie ring that turns orange when you go over, macro bars and meal logging.
- Photo scan: an on-device classifier names the dish, and you adjust the portion before logging.
- Progress: 7, 30 and 90-day ranges, a weight trend line and calorie bars against your goal.
- Settings: your targets, plus clearing today or erasing everything.
Everything is saved on the phone. Screens are in Views and the food data is in FoodDatabase.swift.
Please create a landing page for a cute award winning cafe
What should the cafΓ© feel like?
This sets the colours, the type and the illustrations.
Cute and cosy: soft pinks, honey yellow and chunky outlines- Thought for 5s
- Created index.html +318
- Created style.css +542
- Created script.js +88
- Opened the page with no console errors
I've built Honeybun CafΓ©, a one-page landing site for a cute, award-winning neighbourhood cafΓ©. It's plain HTML, CSS and JavaScript with no build step.
What's on the page, top to bottom:
- Top bar: a menu that stays at the top as you scroll.
- Opening section: the headline "Tiny cafΓ©, big warm hugs.", two buttons and a smiling coffee cup with rising steam.
- Scrolling strip: a tilted band of house favourites.
- Our Story: a short note from the founder.
Files are in index.html, style.css and script.js. Use Publish to put it online.
Why did day 7 retention drop after our 2.3 release?
Which data should I use?
I can query a connected source or read a file you upload.
PostHog, the habit-tracker project that's already connected- Connected to PostHog
- Ran 6 HogQL queries
- Created analysis.py +142
- Wrote report.md +94
Done. Day 7 retention fell from 41% to 36% after 2.3, and nearly all of the drop is new users who never turn on reminders.
What I found:
- Onboarding: 2.3 moved reminders behind a button, and only 38% of new users turn them on, down from 64%.
- Crashes: steady at 99.6% crash-free, so they aren't the cause.
- Next step: put the reminder screen back into onboarding.
The full report is in report.md, with the HogQL to reproduce every number.
All Models overall allowance100% left
Thirsty Models allowance100% left
Resets to 100% tomorrow at 4:00 PM (in 18h 21m)
Context window15%
Files the agent makes in a chat's scratchpad show here.
Thursday, Nov 13
Today3-day streak
Meals
- Nothing logged yet
- π₯£BreakfastOatmeal & berries320
- π₯LunchChicken salad540
- πSnackApple & peanut butter380
- πDinnerPasta primavera610
Your journey
Progress
7 days30 days90 days 0Avg kcal +0.0 kgWeight change 0/7Days loggedYour plan
Settings
- GoalLose weight
- Daily target1,720 kcal
- Protein, carbs, fat130 Β· 180 Β· 57 g
- UnitsMetric
- Clear today
- Erase all data
Tiny cafΓ©, big warm hugs.
Honey lattes, butter-rich bakes and a seat that's always saved for you.
See the menuFind us made withlove β₯
A little note from us
We never set out to win anything. We just wanted to make the cosiest cup in town, with butter-rich bakes, honey in everything, and a seat that's always saved for you. Turns out, a few people noticed.
Retention analysisNov 10, 2026
Why day 7 retention dropped in 2.3
PostHog Β· habit-trackerInstalls Oct 20 β Nov 29,768 new users
Summary
WhatDay 7 retention for new users fell 5.5 points, from 41.2% on 2.2 to 35.7% on 2.3.
Why2.3 moved the reminder prompt out of onboarding. Opt-in fell from 64% to 38%, and people without reminders come back far less.
NextPut the prompt back as the last step of onboarding, and ship it as a two-week A/B test.
1Retention dropped from the first day
The two cohorts split on day 1, 58.0% against 52.3%, and the gap barely moves after that. A drop that shows up on day 1 and then holds points to the first session, not to something that breaks later.
Retention by day since install2.22.3
Figure 1. Share of each cohort that opened the app on each day after install. 2.2: 4,812 people who installed Oct 20β26. 2.3: 4,956 who installed Oct 27 β Nov 2.
2Fewer new users turn on reminders
Opt-in fell overnight on Oct 27, the day 2.3 shipped, and stayed down. 2.2 ends onboarding with a screen that asks for a reminder time. 2.3 replaced it with a βSet a reminderβ button on the Today screen, which most people never tap.
Reminder opt-in by install day2.22.3
Figure 2. Share of new users who turned on a reminder within 24 hours of installing.
3The reminder mix explains the drop
Inside each group, retention barely moved. What changed is the mix: far more people on 2.3 never set a reminder, and they are much less likely to come back. Holding each group at its 2.2 rate, the mix alone accounts for 5.2 of the 5.5 points.
| Day 7 retention | Share of users | 2.2 | 2.3 | Change |
|---|---|---|---|---|
| Reminders on | 64% β 38% | 48.4% | 48.1% | β0.3 |
| Reminders off | 36% β 62% | 28.4% | 28.1% | β0.3 |
| All new users | 4,812 β 4,956 | 41.2% | 35.7% | β5.5 |
4What it isn't
- Crashes. 99.6% of sessions were crash-free on both versions.
- Who installed. Paid installs went from 31% to 30%, and iOS from 72% to 71%.
- Seasonality. The four weekly cohorts before 2.3 all landed between 40.8% and 41.6%.
- Tracking. 2.3 sends the same events, and opens per active user held at 2.6 a day.
Recommendation
Bring the reminder screen back as the last step of onboarding, and ship it as a 50/50 test for two weeks.
+5.2 ptsExpected day 7 retention~260More people retained a week2 weeksTo detect a 3 pt change
Method
- CohortsNew users by first
app_opened, grouped by the version they installed. 2.2: Oct 20β26. 2.3: Oct 27 β Nov 2, 2026. - Day N retentionOpened the app on day N after install, where install day is day 0. Data runs through Nov 9, so everyone has a full week.
- Reminders onFired
reminder_enabledwithin 24 hours of install. - SignificanceTwo-proportion z-test, p < 0.001. The 95% interval for the change is β7.4 to β3.6 points.
- ExcludedInternal and test devices.
Reproduce
Run against the habit-tracker project. new_user_cohorts is the view analysis.py creates.
SELECT version, reminders_on, count() AS users, round(100 * avg(active_day_7), 1) AS day_7_pct FROM new_user_cohorts WHERE installed_at >= '2026-10-20' AND installed_at < '2026-11-03' GROUP BY version, reminders_on ORDER BY version, reminders_on DESC
Written by FreeMilk from 6 HogQL queries Β· Source: PostHog
@@ -0,0 +1,46 @@
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8" />
<title>Dusk Raid</title>
<link rel="stylesheet" href="style.css" />
</head>
<body>
<canvas id="game"></canvas>
<div class="hud">
<span id="kills">KILLS 0</span>
<span id="score">SCORE 0</span>
</div>
<script type="module" src="game.js"></script>
</body>
</html>@@ -0,0 +1,486 @@
import * as THREE from "three";
import { buildCity } from "./world.js";
const canvas = document.querySelector("#game");
const renderer = new THREE.WebGLRenderer({ canvas, antialias: true });
const scene = new THREE.Scene();
scene.fog = new THREE.Fog(0x2b2420, 30, 140);
const camera = new THREE.PerspectiveCamera(70, innerWidth / innerHeight, 0.1, 400);
camera.position.set(0, 1.7, 12);
const state = { ammo: 30, kills: 0, score: 0, health: 100 };
const keys = new Set();
addEventListener("keydown", (event) => keys.add(event.code));
addEventListener("keyup", (event) => keys.delete(event.code));
buildCity(scene);@@ -0,0 +1,266 @@
import * as THREE from "three";
const concrete = new THREE.MeshStandardMaterial({ color: 0x8a7f74, roughness: 0.9 });
export function buildCity(scene) {
for (let i = 0; i < 18; i++) {
const height = 8 + Math.random() * 30;
const tower = new THREE.Mesh(new THREE.BoxGeometry(10, height, 10), concrete);
tower.position.set((i % 6) * 16 - 40, height / 2, Math.floor(i / 6) * -22 - 20);
scene.add(tower);
}
}@@ -0,0 +1,4 @@
.DS_Store xcuserdata/ build/ DerivedData/
@@ -0,0 +1,582 @@
// !$*UTF8*$!
{
archiveVersion = 1;
classes = {
};
objectVersion = 77;
objects = {
/* Begin PBXFileReference section */
0A0000000000000000000003 /* CalorieTracker.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; path = CalorieTracker.app; sourceTree = BUILT_PRODUCTS_DIR; };
/* End PBXFileReference section */@@ -0,0 +1,486 @@
import SwiftUI
struct ProgressView: View {
@Environment(AppStore.self) private var store
@State private var range: DayRange = .week
var body: some View {
ScrollView {
VStack(alignment: .leading, spacing: 16) {
Text("YOUR JOURNEY")
.font(.caption.weight(.semibold))
.foregroundStyle(.secondary)
Text("Progress").font(.largeTitle.bold())
RangePicker(selection: $range)
StatsRow(store: store, range: range)
WeightCard(entries: store.weights)
}
.padding()
}
}
}@@ -0,0 +1,312 @@
import Vision
import UIKit
struct FoodScanner {
func classify(_ image: UIImage) async throws -> [Food] {
guard let cgImage = image.cgImage else { return [] }
let results = try await ClassifyImageRequest().perform(on: cgImage)
return results
.filter { $0.confidence > 0.2 }
.compactMap { FoodDatabase.shared.match($0.identifier) }
}
}@@ -0,0 +1,1940 @@
struct Food: Identifiable, Codable {
let id: String
let name: String
let calories: Int
let protein, carbs, fat: Double
}
let foods: [Food] = [
Food(id: "oatmeal", name: "Oatmeal", calories: 150, protein: 5, carbs: 27, fat: 3),
Food(id: "banana", name: "Banana", calories: 105, protein: 1.3, carbs: 27, fat: 0.4),@@ -1,28 +1,21 @@
import SwiftUI
struct ContentView: View {
- var body: some View {
- Text("Hello, world!")
- .padding()
- }
+ @State private var tab: Tab = .progress
+
+ var body: some View {
+ TabView(selection: $tab) {
+ TodayView().tag(Tab.today)
+ ProgressView().tag(Tab.progress)
+ SettingsView().tag(Tab.settings)
+ }
+ }
}+ 18 more files
@@ -0,0 +1,318 @@
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8" />
<title>Honeybun CafΓ©</title>
<link rel="stylesheet" href="style.css" />
</head>
<body>
<nav class="top">
<a class="logo" href="#">Honeybun <em>cafΓ©</em></a>
<a href="#menu">Menu</a>
<a href="#story">Story</a>
<a class="order" href="#visit">Order now</a>
</nav>
<header class="hero">
<span class="badge">β
Best Neighbourhood CafΓ© 2024</span>
<h1>Tiny cafΓ©, <em>big</em> warm hugs.</h1>@@ -0,0 +1,542 @@
:root {
--cream: #fef3e4;
--ink: #2a1c15;
--pink: #f4a3b2;
--honey: #fde58a;
}
.hero h1 {
font: 600 clamp(2.6rem, 6vw, 4.4rem) / 0.98 "New York", Georgia, serif;
letter-spacing: -0.035em;
}
.btn {
border: 2px solid var(--ink);
border-radius: 999px;
box-shadow: 3px 3px 0 var(--ink);
}@@ -0,0 +1,88 @@
const status = document.querySelector("[data-open]");
const hour = new Date().getHours();
status.textContent = hour >= 7 && hour < 18 ? "Open now" : "Closed now";
const strip = document.querySelector(".strip");
strip.append(...strip.cloneNode(true).childNodes);@@ -0,0 +1,142 @@
from posthog import hogql
COHORTS = {
"2.2": ("2026-10-20", "2026-10-27"),
"2.3": ("2026-10-27", "2026-11-03"),
}
def day7_retention(version):
start, end = COHORTS[version]
people = hogql(
"SELECT person_id, min(timestamp) AS first_seen FROM events "
"WHERE properties.$app_version = {version} "
"AND timestamp >= {start} AND timestamp < {end} GROUP BY person_id",
version=version, start=start, end=end,
)
return retained(people, days=7) / len(people)@@ -0,0 +1,94 @@
# Why day 7 retention dropped in 2.3 > **What:** Day 7 retention for new users fell **5.5 points**, > from 41.2% on 2.2 to 35.7% on 2.3. > **Why:** 2.3 moved the reminder prompt out of onboarding. > Opt-in fell from 64% to 38%. > **Next:** Put the prompt back, as a two-week A/B test. | Metric | 2.2 | 2.3 | Change (pts) | |---|---|---|---| | Day 7 retention | 41.2% | 35.7% | β5.5 | | Reminder opt-in | 64% | 38% | β26 | | Crash-free sessions | 99.6% | 99.6% | 0 | ## 1. Retention dropped from the first day 
# Why day 7 retention dropped in 2.3 > **What:** Day 7 retention for new users fell **5.5 points**, > from 41.2% on 2.2 to 35.7% on 2.3. > **Why:** 2.3 moved the reminder prompt out of onboarding. > Opt-in fell from 64% to 38%. > **Next:** Put the prompt back, as a two-week A/B test. | Metric | 2.2 | 2.3 | Change (pts) | |---|---|---|---| | Day 7 retention | 41.2% | 35.7% | β5.5 | | Reminder opt-in | 64% | 38% | β26 | | Crash-free sessions | 99.6% | 99.6% | 0 | ## 1. Retention dropped from the first day The two cohorts split on day 1, 58.0% against 52.3%, and the gap barely moves after that. A drop that shows up on day 1 and then holds points to the first session.  ## 2. Fewer new users turn on reminders Opt-in fell overnight on Oct 27, the day 2.3 shipped. 2.3 replaced the reminder screen at the end of onboarding with a "Set a reminder" button on the Today screen.  ## 3. The reminder mix explains the drop | Reminders | Share | 2.2 | 2.3 | Change | |---|---|---|---|---| | On | 64% β 38% | 48.4% | 48.1% | β0.3 | | Off | 36% β 62% | 28.4% | 28.1% | β0.3 | | Everyone | 100% | 41.2% | 35.7% | β5.5 | Holding each group at its 2.2 rate, the mix alone accounts for **5.2 of the 5.5 points**. ## 4. What it isn't - **Crashes.** 99.6% crash-free on both versions. - **Who installed.** Paid 31% β 30%, iOS 72% β 71%. - **Seasonality.** The four weekly cohorts before 2.3 landed between 40.8% and 41.6%. - **Tracking.** Same events, 2.6 opens per active user a day. ## Recommendation Bring the reminder screen back as the last step of onboarding, and ship it as a 50/50 test for two weeks: about +5.2 pts day 7 retention, ~260 more people retained a week. ## Method - **Cohorts:** new users by first `app_opened`. 2.2: Oct 20β26. 2.3: Oct 27 β Nov 2, 2026. - **Day N retention:** opened the app on day N after install, where install day is day 0. - **Significance:** two-proportion z-test, p < 0.001, 95% CI β7.4 to β3.6 pts. ## Reproduce ```sql SELECT version, reminders_on, count() AS users, round(100 * avg(active_day_7), 1) AS day_7_pct FROM new_user_cohorts WHERE installed_at >= '2026-10-20' AND installed_at < '2026-11-03' GROUP BY version, reminders_on ```
Your Agent Everywhere
Not just remote control, your agent lives everywhere you want it to work. On the phone, on the Mac, or even on the cloud (coming soon) triggered whenever you need it.
Integrations, MCPs, and APIs
Power your projects and workflows with our native integrations or bring your own.
Stay in Sync with GitHub Support
Local git support as well as a GitHub integration to clone, push, and create PRs on your repos.
Questions, Answered
Something else on your mind? Reach out to us on X.
What is FreeMilk?
FreeMilk is an AI agent for your phone. Describe what you want to build in plain words, and it plans the work, writes the code and builds it.
Is it really free?
Yes, 100%. FreeMilk is supported by ads, so there's no subscription and no per-token bill. You get a daily allowance of models like GLM, Luna and DeepSeek.
Which models can I use?
Your daily allowance covers GLM 5.3 Flash, DeepSeek v4.1 Flash, and GPT-6 Luna. It refills every day, and you can switch between models whenever you like.
Do I need to know how to code?
No. Describe what you want the way you'd explain it to a friend. If you do code, every file is yours to read and change.
What can I build with it?
Small apps, websites, games, scripts, prototypes and tools. If you can describe it, FreeMilk will plan it and build it with you.
Where do my projects live?
On your phone. When you send a prompt, it's processed by the model you picked so it can respond.
Which devices are supported?
FreeMilk is available for iPhone on the App Store.