Apunta

Local-first clinical note drafting for therapists.
Dictate or type the rough version, review the draft, refine it beside the text, and use Finish & copy to put the finished note into the records system you already use.

Read this as a pre-release project. Apunta runs as a desktop app on Linux (an AppImage) with local AI, and its whole test suite runs in fake-AI mode. The macOS build is configured but has never run on a Mac: see docs/v2/MAC-FIRST-RUN.md.

The short version

Apunta is a drafting tool, not a clinical record system. It keeps drafts, transcripts and refinement conversations on your computer, then puts the note on your clipboard when you choose Finish & copy. Read every draft before you use it: a local model can still write a sentence that was never said.

What it feels like

01 · Speak or type the session Record a session summary or write rough notes in any order. Local whisper.cpp transcription shows a live, provisional preview and turns the stopped recording into editable source text.
02 · Review the draft Apunta drafts into your chosen note format. Unclear speech stays marked; sections you did not cover stay blank instead of being filled with plausible fiction.
03 · Refine beside the note Ask for a shorter sentence, a moved section or another concrete edit in the chat beside the draft. The note remains yours to inspect before you finish and copy it.
04 · Keep patients and sessions together Organize notes by patient, with treatment plans, session briefings, backups and an optional per-patient brainstorm kept in the same local practice.
05 · Reuse your note formats Start with the bundled progress-note format, or add your own sections from a blank template, examples or typed names. Change formats later in Settings.

Privacy is a product decision

• The running app talks only to 127.0.0.1 / localhost.
• Writing and transcription use models stored on your computer.
• There is no account, analytics or crash reporting.
• Two network exceptions, both outside the running server and browser app: explicit first-run model acquisition (the installer downloads pinned model files from its allow-list and verifies their checksums), and the desktop app's update check (below).
• Update check (desktop app only). The check contacts github.com and the release-asset host it redirects to. It sends the request metadata any HTTPS request carries (IP address, TLS, user agent, and the target, architecture and version in the URL) and no note content, no patient data and no app-generated identifier. It can be turned off in Settings › About, and the app works offline. (Español: la búsqueda de actualizaciones, solo en la app de escritorio, se conecta con github.com y con el servidor de archivos al que este redirige. Envía los metadatos que lleva cualquier solicitud HTTPS —dirección IP, TLS, agente de usuario y el sistema, la arquitectura y la versión en la URL— y ningún contenido de notas, ningún dato de pacientes ni ningún identificador generado por la app. Se puede desactivar en Ajustes › Acerca de, y la app funciona sin conexión.)
• Apunta has no password of its own. Lock the screen, use a separate user account on a shared computer, and turn on full-disk encryption (FileVault on a Mac, LUKS on Linux) before real notes go in.
• Backups can contain patient data. Keep them encrypted and out of cloud-synced folders.

See docs/INSTALL.md for the plain-language limits and the backup/restore procedure. Settings › About links to this repository, and the licence notices live in THIRD-PARTY-LICENSES.md.

Getting started

For the person using Apunta

On Linux (x86-64), Apunta is a single AppImage:

1. Install Ollama from its website, if it is not there already.

2. Download Apunta_<version>_amd64.AppImage from the latest release.

3. Make it runnable and open it:

   chmod +x Apunta_*_amd64.AppImage && ./Apunta_*_amd64.AppImage

On its first launch Apunta offers to download its speech model and, if Ollama does not have it yet, its writing model; nothing downloads until you press Download. After that it works offline, and it updates itself from the same releases page (signed, and you can turn the check off in Settings). docs/RECOVERY.md walks through setting up a fresh PC for development.

On a Mac, Apunta will download as an Apunta.dmg: open it and drag Apunta to Applications. It carries its own Ollama, and its first launch offers to download the models, so nothing else needs installing. There is no .dmg to download yet. The macOS build is configured (Apple silicon, macOS 14 or newer), and docs/v2/MAC-FIRST-RUN.md is its first build and run, step by step.

For development

git clone https://github.com/villenull/Apunta.git
cd Apunta
npm install
npm run dev:fake

Open http://127.0.0.1:5173. Fake mode keeps the whole app runnable without Ollama, whisper.cpp or any model weights. For the production-shaped local server:

npm start

The source path needs Node 24.19.0. The desktop app wraps the same server in a Tauri shell:

bash scripts/v2/package-linux-resources.sh
npm run tauri:build

Before real notes, run the read-only checks in docs/PREFLIGHT.md.

To reconstruct the exact sanitized Linux reference configuration after a machine reset, including model digests, runtime versions, note instructions and safe verification commands, start with docs/RECOVERY.md. It uses the existing installer and config/recovery/current-linux.json; model acquisition remains an explicit user action.

Development status

Built: local drafting and transcription, note formats, the refine chat, treatment plans, session briefs, a per-patient brainstorm, backups, Halaxy PDF and Claude conversation import, English and Spanish, and the Linux desktop app with a signed updater. After each draft, the server checks what it can verify against the source: it puts back a risk review the draft dropped (in the therapist's own words), removes negative findings about background that was never gathered, rewords diagnostic labels that were never said and drops figures that were corrected. It tells the therapist in the note's chat. Model quality is measured with npm run eval on synthetic fixtures; CI's fake-AI run proves plumbing, not clinical faithfulness.

Open: the first run on a Mac, and real-session use. Read docs/HANDOFF.md for current work and docs/decisions.md for decisions that change the product's boundaries.

Useful commands

Command	Purpose
npm run dev:fake	Vite + Fastify development loop without AI tooling
npm run lint	Lint, format, URL and license checks
npm run typecheck	TypeScript checks across workspaces
npm test	Unit and integration tests
npm run build	Production build
npm run e2e	Playwright against the built app in fake mode
npm run eval -- --fake	Eval harness positive-control self-check
npm run tauri:build	Build the desktop app (AppImage on Linux, .app/.dmg on a Mac)
npm run check:release	Check a release build's identity, update key and endpoint

Repository map

Path	Role
web/	React + Vite browser UI
server/	Fastify API, SQLite and local AI providers
shared/	Zod schemas and types shared by server and web
installer/	First-run model download and checksum verification
src-tauri/	Tauri desktop shell and updater
macos/	The earlier Swift/AppKit shell, superseded by src-tauri/
e2e/	Playwright specs and synthetic evaluation fixtures
docs/	Installation, verification, decisions and handoff

License and status

This repository is a pre-release project and is marked UNLICENSED in package.json: the source is visible, but no permission to redistribute the source or packaged application is granted by this README. Do not put real patient text, audio or exports in issues, fixtures, screenshots or commits.