HiddenSteps

Watches how you work. Spots what you missed. Never lifts a finger.

HiddenSteps is a local-first desktop app that pays attention to your ordinary, repeated work — which apps you jump between, what keeps getting copied from one place to another, which five steps you do every single morning — and tells you, in plain language, what it noticed and what your real options are.

It is not an AI agent. It does not click things, fill in forms, or run scripts on your behalf. It watches, it explains, and then it gets out of the way. What you do with what it finds is entirely up to you — a keyboard shortcut, a five-line script, an n8n flow, an RPA bot, or nothing at all, because sometimes the honest answer is "this isn't worth automating."

Read why it matters if you want the longer version.

• Notice repeated work — the same sequence of apps, actions, or file operations recurring across days or weeks, not just within one session.
• Explain itself — every recommendation states why it was suggested, how confident HiddenSteps is, what it assumed, and what it deliberately ignored. Numbers like "you did this 31 times" always trace back to real observed events, never to something an AI model made up — see how that's enforced.
• Span the whole solution space — a suggestion can be as small as "there's a keyboard shortcut for that" or as involved as "here's a hybrid script + local-LLM approach," with the tradeoffs of each laid out side by side.
• Run entirely offline — local AI models (via Ollama and friends) by default; cloud providers are opt-in, per-provider, and never see anything above the privacy level you've chosen.
• Show you exactly what it's watching, in real time — a live feed of what was just captured (after redaction), one-click pause, one-click delete-everything, and an audit log of every privacy-relevant change it's made.
• Redact before it remembers anything — passwords, API keys, credit card numbers, and other secret-shaped text are stripped before anything touches disk; if it's not sure, it drops the observation entirely rather than guess.

• Record your screen or keystrokes by default.
• Take any action without you explicitly approving it, every time.
• Send anything to the cloud unless you've turned that on, provider by provider.
• Report anything to an employer, IT admin, or anyone but you — there is no "manager view," and there never will be one. See why.
• Require an account, a subscription, or an internet connection to work.

Full detail, including exactly what's retained and what could ever leave your device: docs/design/05-privacy-model.md.

Grab the latest build for your OS from Releases:

Launch it, and the first-run flow walks you through exactly what it does, what it won't do, and lets you pick a privacy level and an AI provider before anything is observed.

This is an early, unsigned, prerelease build — see CHANGELOG.md for what's in it and what's known to still be missing. If you want the paper trail behind it — competitive research, architecture decisions, UX wireframes, the implementation roadmap — it's all in docs/, and every claim in this README traces back to something in there.

Prefer to build from source, or want to contribute? Two independent pieces:

git clone https://github.com/JGalego/HiddenSteps.git
cd HiddenSteps
cargo test --workspace       # 183 tests with no system dependencies beyond cargo (+10 in hiddensteps-observation that need a live X11 display)

If you have Ollama running locally, you can also run the tests that hit a real model:

cargo test -p hiddensteps-llm-provider -- --ignored

cd apps/desktop/ui && npm install && npm test   # 50 tests, real jsdom rendering

The native shell needs Tauri's platform prerequisites (on Linux: webkit2gtk-4.1 and friends via your package manager) — verified building on all three OSes in CI:

cd apps/desktop/src-tauri && cargo build
# once it builds: `cargo tauri dev` from apps/desktop/, with the ui/ dev server running

Tauri (Rust core + native OS WebView) over a modular Rust workspace, one crate per responsibility — observation, redaction, the event pipeline, pattern detection, the recommendation engine, LLM providers, the privacy-dispatch gate, a WASM plugin sandbox, enterprise policy — all behind Clean Architecture boundaries, with a single SQLCipher-encrypted file as the only place anything durable lives. Every non-obvious choice has a written rationale: see the Architecture Decision Records.

docs/
  research/   Competitive analysis, market gaps, risk/ethical/privacy analysis, threat model
  design/     PRD, ADRs, system architecture, data flow, trust/privacy/security models, DB schema, plugin & API specs
  ux/         User journeys, onboarding wireframes, privacy dashboard, recommendations UX, settings IA, accessibility
  roadmap/    Milestones, technology choices, testing strategy, security/privacy/performance test plans
crates/       The Rust core — see crates/README.md for exactly what's built and how it's verified
apps/desktop/ The Tauri shell + React/TypeScript UI — see apps/desktop/README.md
.github/workflows/ci.yml       Cross-platform build/test on every push (Linux, macOS, Windows)
.github/workflows/release.yml  Builds and publishes installers to GitHub Releases on a version tag

Genuinely welcome, especially: actually running a release build day-to-day and reporting back what breaks; a real design pass on the app icon; signing/notarization setup (there is none yet — see the release notes); and filling in any of the disclosed gaps listed throughout crates/README.md and apps/desktop/*/README.md.

See CONTRIBUTING.md for setup, the checks to run before opening a PR, and the ground rules; report security issues privately per SECURITY.md.

MIT.