LockIn

A personal study & resource hub — not an AI generator. You're the brain; LockIn is the shelf + planner. Everything is organized around Subjects.

• Subjects (e.g. "Operating Systems", "GATE Prep") — each holds a plan and resources.
• Plan — an ordered list of Milestones (phases) with markdown notes, each holding a checklist of Tasks.
• Resources — saved URLs only: web links, AI chat links, book/PDF references.
• Today — a cross-subject view of tasks due or overdue.
• Focus — a timer (Pomodoro / stopwatch) that logs time against a task.
• Progress — streaks, completions, focus time and a GitHub-style activity heatmap.
• Ask (⌘J) — chat with an LLM running on your own machine or entirely in your browser, with your subjects, plans and notes as context — retrieved rather than blindly trimmed once a plan outgrows one prompt.

Single user per account. Two interchangeable looks via the theme toggle: Creative (light, neobrutalist) and Focus (dark, editor-calm).

Stack

Next.js (App Router) · React 19 · TypeScript · Prisma + PostgreSQL · NextAuth v4 (JWT) · TanStack React Query · shadcn/ui + Tailwind v4 · next-themes.

Getting started

npm install
npm run dev        # http://localhost:3000 (auto-bumps to 3001 if taken)

Set the following in .env (Prisma CLI) and .env.local (the app):

DATABASE_URL=            # Postgres, runtime. Supabase: transaction pooler (:6543, pgbouncer=true)
DIRECT_URL=              # Postgres, migrations. Supabase: session pooler (:5432, NO pgbouncer params)
AUTH_SECRET=
NEXTAUTH_SECRET=
NEXTAUTH_URL=            # e.g. http://localhost:3001
GOOGLE_CLIENT_ID=
GOOGLE_CLIENT_SECRET=

DIRECT_URL must not be a copy of DATABASE_URL: pgbouncer in transaction mode can't run DDL, and prisma db push hangs forever instead of erroring.

AUTH_SECRET is not optional. NextAuth's own routes fall back to an internal secret, so sign-in appears to work without it — but every API route reads the JWT via getToken and returns 401, so the app loads and shows no data.

Commands

npm run dev      # Next.js + Turbopack
npm run build    # prisma generate && next build
npm run start    # serve the production build
npm run lint     # next lint

npx prisma db push   # sync schema to the database
npx prisma studio    # browse data

Data model

User ──< Subject ──< Milestone ──< Task ──< TimerSession, and Subject ──< Resource.

The Prisma client is generated into app/generated/prisma/ (committed) — regenerate with npx prisma generate after schema changes; never hand-edit.

Ask — local LLM chat, in-browser or your own server

Press ⌘J anywhere to open the chat panel. On a subject page it starts scoped to that subject, so questions are answered against your real milestones, tasks and notes; deselect everything to search everything, or for a plain chat.

The chat panel's settings offer two providers:

• Local server — the model runs on your machine and the browser talks to it directly; LockIn's server assembles the prompt and stores the transcript, but never sees your endpoint, model, or API key. Any OpenAI-compatible server works: Ollama, LM Studio, llama.cpp, Jan, vLLM. Set the server URL (presets provided) and hit Test connection to list models.
• In browser — a model runs entirely on your GPU via WebLLM, no install and no CORS setup. Pick a model from the shortlist (sizes shown) and hit Download & load; weights are cached by the browser, so later visits skip the download. Needs a WebGPU-capable browser (recent Chrome or Edge) — Safari and older machines fall back to the local-server path.

Because the local-server path calls your model from the browser, that server has to allow the page's origin:

OLLAMA_ORIGINS=https://your-lockin-domain.com ollama serve

# macOS app
launchctl setenv OLLAMA_ORIGINS "https://your-lockin-domain.com"   # then restart Ollama

Running both on localhost needs no configuration. LM Studio has a CORS toggle in its server tab.

Retrieval

Once a plan grows past what fits in one prompt, LockIn stops trying to trim it and starts searching it. Notes are chunked and embedded in the browser (WebGPU, via WebLLM's fixed embedding model) the first time the panel opens, then re-embedded incrementally as they change — the status row above the composer shows progress and offers a rebuild. A question then gets a full outline of your plan (subjects, milestones, open tasks) plus the handful of note passages that actually score as relevant, instead of a blind cut. This works independently of which chat provider you've picked, including the local-server path — it only needs WebGPU for the embedder. Small plans are unaffected: if everything already fits, the full plan is sent as before. Toggle it off in settings ("Search notes for context") to always fall back to the old trim-and-send behaviour.

Assistant replies that used retrieval show a collapsed sources row listing which passages they drew on — useful for telling "the model reasoned badly" apart from "the model never saw the right note."

Assistant replies carry a save to note action that appends the answer to any milestone, task or subtask note. Nothing else in your plan is ever written by the model.