CoreOS

Capture first. Organize later.

A local-first personal second brain for free-form text capture, persisted on device and organized automatically by an AI classifier.

repository · ezequielmenor.es

English · Español

---

What is CoreOS?

Most note and task apps make you decide where something belongs before you can save it: pick a notebook, pick a project, pick a list, pick a tag. That decision cost is what stops most captures from happening.

CoreOS removes it:

Capture → Persist locally → Classify with AI → Organize automatically

You write free-form text. It is stored on the device before any AI call. An LLM then classifies it into a structured table. The raw input is never lost, never rewritten, and organization happens after capture, never before.

Screenshots

Today	Capture	Library
Today's tasks and pending captures at a glance.	One field. Anything you want to remember.	Notes grouped by recency, with full-text search.

How it works

flowchart LR
    A[Free-form input] --> B[SQLite inbox]
    B --> C[LLM classifier]
    C --> D[Dispatch by type]
    D --> E[Notes]
    D --> F[Tasks]
    D --> G[Expenses]
    D --> H[Habits]
    D --> I[Sleep]

Loading

Key properties:

• Raw input is persisted first — the SQLite insert happens before the network call.
• raw_text is canonical — for notes it is stored verbatim as body_md. The LLM only emits metadata (suggested title and tags); it never rewrites, summarizes or splits the user's text.
• AI failure never discards captures — an invalid or failed model response leaves the inbox row in pending status, retryable from Today.

🔒 Local-first by design

• SQLite database lives on the device
• No backend, no remote service
• No account, no auth
• API keys held in expo-secure-store (iOS Keychain / Android EncryptedSharedPreferences)
• Capture persisted before any AI call
• AI failure leaves the capture pending — raw text is never lost

Current V1

Three tabs (Today, Capture, Library) plus Tasks and Settings as secondary screens.

Today — pending and overdue tasks, sorted overdue-first then by priority. Inline complete. A chip shows unclassified captures and retries the pipeline on tap.

Capture — a single free-text field. On save the text is inserted into the local inbox and the async classifier runs. Also reachable from the iOS share sheet and the coreos://capture deeplink.

Library — notes grouped into Pinned / Today / Yesterday / This week / Earlier. Full-text search, tag filters, Markdown editor with autosave, pin, soft-delete with undo and restore.

Tasks — full CRUD with date normalization for free-form due_date inputs (hoy, mañana, d/m/y, ISO → YYYY-MM-DD).

Architecture details

Pipeline invariants

Documented in src/services/inbox.ts and enforced in code:

• The inbox row is marked processed inside the same transaction as the dispatch, guarded by WHERE status='pending' (lock optimistic).
• Items are processed sequentially (for/await, never Promise.all).
• Exported pipeline functions never throw — they return a result the caller can retry.
• A module-level mutex (_batchInFlight) prevents concurrent batches; late captures automatically trigger one extra drain pass.

Source layout

src/app/         expo-router screens (file-based, typed routes)
src/stores/      Zustand stores
src/db/          SQLite singleton, migrations, dispatcher
src/db/queries/  per-domain query modules
src/services/    LLM client + inbox orchestration
src/components/  shared UI
src/hooks/       theme, note editor
src/lib/         animations + note save gate
src/constants/   theme tokens

Notes schema & search

• Notes have a FTS5 virtual table (notes_fts) with tokenize='porter unicode61', ranked with bm25.
• Triggers keep the FTS index in sync on INSERT / UPDATE / DELETE.
• Soft-delete via deleted_at; restore sets it back to NULL.

Classification types

The LLM emits one of five types: nota, gasto, tarea, habito, sueno. Schema for each is enforced client-side in processInboxText; if the response is missing required fields or returns an unknown type, the pipeline logs a warning and leaves the row in pending so the next batch re-picks it. The raw_text is never lost — the row only flips to processed once a valid dispatch succeeds inside the same transaction.

Native tabs

expo-router's unstable-native-tabs render the iOS native tab bar on iOS; Android and web fall back to a JS implementation. This is why a dev build is required (and Expo Go is not enough).

Tech stack

Layer	Choice
Framework	Expo SDK 57, React Native 0.86
Language	TypeScript, strict
Routing	expo-router (file-based, typed routes)
State	zustand
Database	expo-sqlite (FTS5)
Secrets	expo-secure-store
Animation	react-native-reanimated 4 + worklets
AI	Any OpenAI-compatible chat completions endpoint

Status

v0.1.0 · Personal V1 · Active development

Not a commercial product. Some tables in the pipeline (gastos, habitos_log, sueno_log) are already written by the classifier, but their management UI is intentionally not part of V1 — the data is captured and stored; the screens come later. There is no automated test suite yet.

Running locally

npm install
npx expo run:ios

A development build is required: the project uses native config plugins (FTS5-enabled SQLite, SecureStore, iOS share extension) and native tabs, so Expo Go is not a supported target.

Open Settings inside the app and configure the AI provider (base URL, API key, model). Keys live in expo-secure-store; there is no .env and nothing sensitive is committed. Without a key, captures are still saved to the inbox and stay pending until one is configured and the pipeline is retried.

npm run lint      # ESLint via expo lint
npx tsc --noEmit  # typecheck

expo-secure-store is unavailable on web — the AI pipeline does not work in a browser build.

Out of scope / not in V1

• Backend
• Cloud sync
• Multi-user / accounts
• Hermes vs JSC engine work
• Graph view and embeddings (note_embeddings exists in the schema but is unused)
• Expense / habit / sleep management UI

Author

Built by Ezequiel Menor — ezequielmenor.es · github.com/EzequielMenor