🚀 Significant update to Hash, a Privacy-First AI Journaling App for Bitcoiners (and nocoiners too).

Quick aside: yes, this NOSTR note was composed on Hash.pink and published from there. The NOSTR integration works great!!! This is a REPOST from my dev environment, essentially my stripHTML function got a lil over-excited and was stripping out elements that I wanted to retain. Hopefully this one posts with the correct line breaks.

Let's talk about the recent update and orient you to Hash, if you haven't heard me talk about it before.

Most AI journaling apps send your thoughts directly to OpenAI/Anthropic. Your private reflections become training data.

Not Hash.

Our Privacy Gateway (released 1 hour ago) sits between you and LLMs:

✅ Strips PII before transmission (emails, names, SSNs, etc—gone)

✅ Rotating pseudonymous IDs every 6 hours (no cross-session tracking)

✅ 45% token reduction through intelligent preprocessing

✅ Local caching prevents redundant LLM calls

The Privacy Gateway pre/post processors are what make Hash special. The LLM is just the "patty"—you can swap between beef, chicken, or pork, but you're still getting a tasty sandwich! I don't know if this is truly the first time private interactions have been available with OpenAI/Anthropic/etc, but it is certainly one of the first products to offer this seamlessly (and for free to all users). So, yes, you can use whatever model you already use, except you can now do it more privately.

Hash is currently LLM agnostic and the privacy layer runs locally on my servers for now, but soon you'll be able to host the pre/post processors on your own device, no technical skills required, you just choose where to store the programs and they install and will work offline, for the most part. It's AI built by a deranged & paranoid freedom maxi who has also designed consumer-facing products and signficant business tools for Fortune 10 clients, startups, and the Department of Defense for 10+ years.

Hash offers:

4 default models (Mixtral 8x7B, Llama 3.1 70B, Claude 3.5 Sonnet, GPT-4o) to start with

YOU can request (almost) ANY model with justification and I'll get it integrated in short order at no cost to you, then it's just a couple clicks for you to switch models AND all of your context is preserved for the model, if you want to enable sharing with it (coming soon)

Post-21-message surveys validate new models for community

Model Cage Match (Ellipsis tier exclusive): Send your prompt to multiple models simultaneously and choose the best one for your task.

And here's where it gets cool: Full transparency mode shows you:

Exactly what PII was stripped

Token reduction percentage (with progress bar!)

Cost savings visualization

Processing pipeline breakdown

This is an insane feature that not only gives you better quality and increased privacy, it also saves you money. IMHO, it should be the standard way AI works, but I think this more of an El Salvador/Bitcoin expat viewpoint.

Pricing:

Free Trial: Test drive with limited credits—no card required

Comma (,) - $7/mo: 50k credits (~500k tokens), 60 min voice transcription

Period (.) - $21.00/mo: 200k credits (~2M tokens), 180 min voice, and more

Ellipsis (...) - $210/mo: 1M credits (~10M tokens), unlimited voice, up to 4 team seats, Caret^ included for account holder (formerly known as "all access pass," coming in 2026)

$5 Credit Top-Ups - Buy 35,000 banked credits that never expire. Subscribers only. Confirmation required before charging (we believe in friction for financial decisions).

15-17% annual discount on all tiers with some easter egg pricing levels in there.

Try it: https://hash.pink!

And if you're wondering if there's more, yes, there is more. Here's the high level technical details + a few other things in the release from tonight.

This is the technical problem I've been trying to solve: how do you use powerful LLMs without compromising user privacy?

Here's the architecture that makes it work:

Step 0: facilitiate nsec login with NOSTR credentials, email not required

Step 1: PII Stripping with regex patterns

Step 2: Pseudonymization, every user gets a rotating hash ID (6-hour rotation)

This means:

❌ LLM providers can't track you across sessions

❌ No user profiling

❌ No cross-session pattern detection

❌ Increased insulation from Chrome/etc sharing your profile information with third-party apps

Your privacy isn't a promise—it's enforced by design.

Step 3: Context Compression

Target: 45%+ token reduction

How?

Remove filler words

Compress repeated concepts

Optimize prompt structure

Maintain semantic meaning

Result: Same quality responses, way less cost. Like an actual free economy, prices should go down over time as the model self-teaches to be more efficient and accurate and higher quality. Since the goal is to help users run this locally, it reduces costs further over time for users.

Step 4: LLM Agnosticism

The Privacy Gateway works with ANY model:

Mixtral 8x7B (fast, cheap)

Llama 3.1 70B (balanced)

Claude 3.5 Sonnet (reasoning)

GPT-4o (flagship)

Users can REQUEST new models. We validate through usage surveys. Some models use more/less tokens, so it won't be consistent, but you will have actual model freedom and actual freedom from models.

Model Cage Match feature: Send ONE prompt → get responses from ALL 4 models. This is a cool feature, but it's really just submitting the prompt to multiple APIs and then presenting the results in a pleasing UI. Still...

Transparency mode shows:

Token reduction percentage

PII stripped (what changed)

Cost savings

Processing time per model

Education = verified trust

Why this matters for Bitcoin/NOSTR folks:

Privacy is non-negotiable, your thoughts stay yours

Open architecture, no vendor lock-in

Cost transparency, see exactly what you're paying for

Community-driven, users shape the model selection

Hash also now has "Easy Reader" mode which makes the text an appropriate width so it's easier to read and write on any device. The Subscription page (when logged in) is beautiful and *should* work, but it might have some hiccups--bear with me! You can also manage your subscription, upgrade/downgrade/cancel with a couple clicks, and on the Ellipsis tier you'll be able to gift Hash to three other folks. For Ellipsis, I haven't thoroughly tested this yet, so if you want to buy it just send me a DM and I'll double check that everything is working properly before you part with your dollars. If you buy it before I've gone back through the testing and squashed bugs, I'll figure out how to make it right, and then I will make it right. Easy enough, right?

The Bitcoin integration will be live soon too, I am currently debating just doing a Lightning invoice or whether I want to integrate a full LSP into my apps. The reason I waited so long on setting up payments is I really want them to be managed centrally for the Caret^ tier since I want users to have one payment to manage instead of having to sign up on multiple sites, but honestly it's time for these apps to start making a little cash and paying off the debt/sweat I've put into making them great.

I am so proud of this app. On a personal level, Hash started off as an afternoon project in late July or early August (not positive) to test out some of the product integrations I've been working on for the last year. I liked it, so I kept working on it. Now, 4 months later, I feel confident that it's one of the best journaling apps you can use. Since I'm able to build fast, cheap, and high-quality with my ecosystem tools, I can ship new features faster than most teams of developers and even many companies. I'm responsive. If you subscribe to the Ellipsis tier, I'll give you my personal email/cell and you can call me 1-2 times a month, talk about AI or Bitcoin or whatever you want, and you can ask me for features and new apps, integration help, or just talk about life. I'm building FOR YOU because without you, I can't build anything. I need you to use my apps so that we can build a brighter future together.

So, use Hash if you're thinking about New Year's Resolutions and want to journal, or you want to have a better way to store/track photos and document family vacations for the memories. Gift Hash to an older parent/grandparent who wants to record their life's history--in the next couple of months, I'll be developing an auto-novel approach within Semi (my brain/processor AI agent) where just by talking and writing a user can generate high quality, book-length transcripts. If you've got younger kids, reach out and tell me what your kids need help with in school... I'll build a Pipe for your family, you can install it in Hash, and once SemiSchool is working better, I can help your kids value research and learning in a way no other AI can because my models aren't focused on giving you all the answers, they're aimed at helping you learn how to think more effectively. Your kids will then have a permanent record of all their learning, so applying to college and preparing for jobs is simple... it's all in one place, automation-enabled. To put it simply: I aim to change human history, nothing less, because I want to restore human agency to AI and next-gen technology.

I'm super excited about all this stuff and can't wait to hear what you think! Let me know if you run into problems, I'm responsive and will usually ship a fix within 24 hours. How many of your other software vendors/programs will give you the phone number of the CEO, CTO, and developer and tell you to call any time you need something? :D

LFG!