Privacy — Human Layer

The Short Version

Human Layer is built so that the most honest thing you can write has a home that can't read it. Your writing lives encrypted on your Mac. The AI that reflects on it runs on your Mac. After a one-time model download on first run, it works without internet. There's no account, no analytics, no telemetry.

What We Don't Collect

No account. You don't sign up. There's nothing for us to associate your writing with.
No analytics, no telemetry. No page-view tracking on this site, no usage events from the app. We don't know how many entries you've written, when you opened the app, or which features you use.
No third-party scripts. This site loads two fonts from Google Fonts (so typography renders) and nothing else. The app makes no third-party requests at all.
No cookies. No advertising IDs, no fingerprinting, no cross-site tracking.
No "anonymized" aggregate data. We're not collecting word counts or sentiment scores or topic distributions in the background.

The honest exception: when you visit this site, our hosting provider (Cloudflare) sees the HTTP request — IP, user agent, timing — the same way any web server does. That's normal log data and Cloudflare's policy governs how long it's kept.

What Stays on Your Mac

Everything you actually write, plus everything Human Layer derives from it.

Journal entries — encrypted with XChaCha20-Poly1305 under a key derived from your passphrase via Argon2id. Live in ~/HumanLayer/.
Reflections from Reading, Ask, Patterns, Mirror, and Signal — generated by an on-device model that runs on your Mac. Never leave.
Embedding vectors for semantic search — built in memory at unlock, wiped on lock.
Your passphrase — we don't have it. There is no back door.

If You Lose Your Passcode

We can't help you. There is no recovery email, no account-side backup, no support process that ends with you getting your writing back. The data is encrypted with a key only you hold.

This is the price of the privacy promise — and it's the right trade. Pick a passphrase you'll remember, write it down somewhere physical if you must, and consider exporting an encrypted backup you store outside the app.

How Reflections Work

Reading, Ask, Patterns, Mirror, and Signal all run on a small on-device model (Qwen3-4B) that lives on your Mac, downloaded once on first run. No external AI provider is contacted. No prompt, no passage, no output ever leaves the device.

When you ask your vault a question — or when a reflection surfaces a pattern — the same engine fires: your query is embedded locally with a small on-device embedding model (nomic-embed-text), the top semantically-similar passages are retrieved from your in-memory index, and those passages plus the query plus a surface-specific prompt are passed to the local chat model. The model answers with inline citations back to the source passages so every claim is traceable to your own words.

Email to Vault

This is optional, and off until you set it up. Incoming mail to inbox@thehumanlayer.co arrives at our edge the way any email does — in the clear, because ordinary email is not end-to-end encrypted. The edge immediately seals it to your vault's public key (a libsodium sealed box) and stores only the sealed blob; from that moment only your laptop, with the matching private key, can read it. After your laptop pulls a message, the stored copy is deleted.

The Calls the App Makes On Its Own

Your writing never leaves your Mac. The app reaches the network in exactly three ways, and it ships a machine-readable Internet Access Policy (the kind a tool like Little Snitch reads) that declares them:

Version check. On launch, one request to api.thehumanlayer.co/updates/latest.json asking what's the current version? No identifying data, no telemetry.
Model download. On first run only, the on-device AI model (about 3 GB) downloads from models.thehumanlayer.co. A download only; nothing is uploaded.
Email-to-vault. Optional. Only after you set up and verify an inbox does the app poll your own mailbox for new messages while it is open.

Under the Hood

The short, technical facts behind the privacy claims. Each one is verifiable from the open source.

Open Source

The code that holds your words — encryption, the vault format, and sync — is public at github.com/joey9prints/hl-core. The app is proprietary; the part you have to trust is not. Read it, run it, and check these claims for yourself.

Key Derivation

Argon2id from your passphrase (m=64 MiB, t=3, p=4) — the modern memory-hard KDF, resistant to GPU/ASIC brute force.

Encryption

XChaCha20-Poly1305 AEAD on every entry. 192-bit nonce space lets us use random nonces safely; constant-time decode resists timing attacks.

Vault Location

~/HumanLayer/. Run xxd on any entry file — you'll see ciphertext. No back door.

AI Runtime

An on-device model (Qwen3-4B) that runs on your Mac, downloaded once on first run. No external API is ever contacted for reflections.

Backups

HLBKUPv1 — a single encrypted archive using the same primitives as the live vault. Portable across machines you own; readable only with your passphrase.

Deeper

Local Intelligence →
Semantic search over your writing. How it stays local, what gets wiped on lock, what swap-to-disk does, and why FileVault is the right companion.

Questions, Mistakes

If you find something on this page that doesn't match what the code or the app actually does, that's a bug we want to hear about. Email joe@thehumanlayer.co and we'll either fix the wording, fix the code, or both.

For the technically curious: the cryptographic core that governs these guarantees — encryption, the vault format, and sync — is open source at github.com/joey9prints/hl-core. The app itself is proprietary; the part you have to trust to keep your words private is the part that is public. That the on-device model never phones out is verifiable a different way — watch the network: the app ships an Internet Access Policy declaring exactly the calls it makes.