Coach.
I built an AI accountability coach for myself, and just open-sourced it at github.com/allurco/coach.
It started as a financial coach — that was my problem at the time — but the framework is goal-agnostic. Anywhere the missing piece is a system that doesn't get tired of asking "did you actually do that today?", this fits: getting out of debt, training for something, learning a language, shipping a side project, breaking a habit.
Why
A few months ago I had a structural deficit. Income was fine, the math wasn't. I needed someone — or something — that wouldn't get tired of asking "did you actually pay that today?" every morning.
A human consultant felt heavy: scheduling, emotional load, the cost. A note-taking app wouldn't push back. So I started building a thing that does both: keeps the plan, and pings me about it.
The first version was just a Laravel page with a Gemini-powered chat. It got useful fast enough that I kept adding to it — PDF analysis of fatura/extrato, action tracking with deadlines, scheduled email pings, and a memory layer so it remembered what we'd already decided.
After a few weeks of using it daily, I figured anyone in a similar mess might want the same tool. So I rewrote it for multi-tenancy and made the repo public.
What it actually does
Open the app, type in plain Portuguese (or English) what's going on. The Coach:
Interviews you on first contact — figures out the income/expense picture, identifies traps (cheque especial, parcelamento at 491% APR), proposes 2-3 concrete actions for the week.
Reads PDFs you upload — fatura, extrato, boleto. Gives you a structured table + a qualitative analysis (e.g. "I see R$ 762 in Google Cloud charges on your personal card; that should probably move to your business card").
Tracks your plan — actions with deadlines, priorities, categories. You mark them done as you go. The Coach pings about what's stuck.
Sends a morning brief at 8am ("Foco do dia: pay the smaller card before the bigger one fires interest"), a weekly recap on Sunday evening, and a stuck-action nudge mid-week.
Replies via email work — answer the morning brief from your phone and the Coach updates the plan. The reply threads back to the same conversation via subaddressing on the Reply-To.
Remembers across conversations — long-term memory of facts you consolidated weeks ago. So when you say "the Santander fatura from May", it knows what you mean without you re-uploading.
Why open source
Two reasons.
One, financial trouble is way more common than people admit, and the apps in this space are either banking-aggregator surveillance (your data → ad targeting) or sterile budget calculators that miss the human part. Someone running their own Coach on their own infra keeps their data private and gets a tool that's actually opinionated.
Two, I'm using this thing every day. Open-sourcing it forces me to keep it in a state where someone else could actually clone and run it — which is the bar I want for software I rely on.
Stack
Laravel 13 + PHP 8.4 for the backend
Filament v5 for the admin/chat UI (chat-first, no dashboard)
Laravel AI SDK with Gemini 2.5 Flash as the model
Livewire 4 streaming so the Coach types in real time
Resend for outgoing email + inbound webhook
Pest for tests (75 right now, including multi-tenant isolation)
SQLite local, MySQL in production
Tailwind v4 + custom CSS for the conversation experience
The whole thing is self-hosted. There's no SaaS dependency beyond the Gemini API key (free tier covers a single user) and Resend (free tier covers a single domain).
Multi-tenant from day one
When I rewrote for sharing, I went row-level: every action and every memory belongs to a user_id, with global scopes ensuring cross-tenant access is impossible. There's no public registration — admins invite users from the avatar menu, and each invitee sets their own password via a one-time link. Closed-loop.
How to try it
git That's it locally. The seeder prints a one-time admin password. Open the app, log in, start typing about your situation. The Coach goes into onboarding mode, interviews you, and starts building your plan from scratch.
For production deploy on Forge or Cloud, the README walks through DNS for Resend (the gotcha is don't put receiving MX on a domain already running Google Workspace), Cloudflare SSL settings (use Full, not Flexible — Flexible breaks session cookies), and the PHP 8.4 requirement.
If you build something on top of it, or fix a bug, PR away. The repo is at github.com/allurco/coach. MIT license.