By default, home page gives all models in the leaderboard, local and hosted. Search for models in the search box on the home page to find the top models by ranking, local(by size) and hosted (by price).

You can also do deep querying/sorting/searching filters of models in each of these three nodes (see the other tabs on top).

The next steps I am working on (would love feedback on this or anything else):

Phase 1: - Change clicks on home page model tiles in one column to search and show models filtered by that across Artificial Analysis, Ollama, OpenRouter - User specifies their system VRAM (unified/dedicated) and we automatically filter the home page with models that would fit on that RAM - in the three columns. - User specifies their price range (per MTok, max across input and output), and we similarly filter and rank by those models across all columns. - User specifies both (VRAM and price range), and we filter by both - leaderboard is union of local and hosted, local by VRAM and hosted by price range match.

Phase 2: Once I have this working, add a local desktop client that automatically reads user system and infers VRAM, renders app as webview. Considering pyside6 with Qt for this.

Phase 3: On desktop client, user can download and chat with the local models automatically based on leaderboard, optionally call hosted models, etc. Used primarily to evaluate and compare local vs hosted models for user's use cases. Also have some interesting alternate experiences to host within the local private app for user to interact with llms, agents, etc.

Do let me know whether this seems useful, or how I can make it more useful.

- sorting by most recent - autoupdating as new models are loaded (every hour).

Comments URL: https://news.ycombinator.com/item?id=48456025

Points: 2

# Comments: 2

This is a scratch my own itch kind of app.

The best way to get the most out of Claude Code is to learn from past sessions. Review past sessions, see where it worked, where it went wrong and what we could have done differently. However, the sessions are saved across multiple json files with cross references, and are hard to read. In a live session, many steps are hidden by default and cannot be easily accessed later. And the mindset when grappling with Claude Code in a session is not exactly ripe to later remembering the details of what went well or wrong and why. So, I built this viewer (using Claude Code, of course) to show Claude Code chats in a single place.

Features:

See detailed steps of Claude interactions - thinking, tool calls/responses, code gen, code diffs. Search, sort, filter. Export individual sessions, repos or entire history as markdown. 100% local and private single page HTML app that reads your claude code chats.

Usage:

Just click the link, use code SHOWHNFREE to download the html page for free, and open it in any Chromium based browser like Google Chrome or Edge (Note: Safari and Firefox won't work!).

Then, open the folder where Claude code saves its sessions (by default, ~/.claude on Mac). Enjoy!

Feedback:

Would love any feedback or suggestions - but hey - it is a single page html - so tweak it as you like - or even better, just ask Claude Code to do it :-)

This is a scratch my own itch kind of app.

The best way to get the most out of Claude Code is to learn from past sessions. Review past sessions, see where it worked, where it went wrong and what we could have done differently. However, the sessions are saved across multiple json files with cross references, and are hard to read. In a live session, many steps are hidden by default and cannot be easily accessed later. And the mindset when grappling with Claude Code in a session is not exactly ripe to later remembering the details of what went well or wrong and why. So, I built this viewer (using Claude Code, of course) to show Claude Code chats in a single place.

Features:

See detailed steps of Claude interactions - thinking, tool calls/responses, code gen, code diffs. Search, sort, filter. Export individual sessions, repos or entire history as markdown. 100% local and private single page HTML app that reads your claude code chats.

Usage:

Just click the link, download the html page for free (use code SHOWHNFREE), and open it in any Chromium based browser like Google Chrome or Edge (Note: Safari and Firefox won't work!).

Then, open the folder where Claude code saves its sessions (by default, ~/.claude on Mac). Enjoy!

Feedback:

Would love any feedback or suggestions - but hey - it is a single page html - so tweak it as you like - or even better, just ask Claude Code to do it :-)

Comments URL: https://news.ycombinator.com/item?id=46545981

Points: 1

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=46388508

Points: 1

# Comments: 1

But if the task is long running, I like to have multiple AI chats - from same or different projects open - so when one is working, another can be nudged forward. I found it helps to pick unrelated things to do in this mode - if two chats are on very closely similar projects or features, it can get confusing and harder to maintain each context in the mind.

I also sometimes pick some (relatively light kind of learning project and do that when I wait for AI to finish something.

This may seem strange right now, but very soon, we will see that intelligent (AI) systems in various companies (OS, browser, cloud, AI labs) are learning much more about us and using that to maximize their own agendas - for example, these companies' AI systems can start recommending products I am likely to buy as part of my future interactions with them. Not happening broadly yet, but only a matter of time. And local private experiences will be very attractive then.

Next: Why Mac app? Currently, Macs with Apple Silicon (M1, etc. processors) can run the largest local models efficiently (compared to other alternatives). And Apple is also more committed to privacy of user data than others. Hence, mac apps. Would also consider iOS apps, but the app store review is an unpredictable bottleneck - so will only do it if experience is validated on Mac.

You asked how it is working on this: I have been learning to use AI assisted coding more efficiently, finding the balance between manual and vibe coding - reached critical mass recently - built and launched my first such mac app in 10 days from idea to launch - using AI assistance more for some parts, less for some. Now, it seems like if I can prioritize what to build (there are too many ideas possible, and most are claimed to be solved by several companies when you search - finding one that can work as an indie dev is the biggest challenge), there can be a pipeline of several small useful apps on this stack.

I have tried many alternatives for the tech stack to build the mac app - swift, tauri, electron, beeware toga, kivy, etc., and finally settled on pyside 6 (a more leniently licensed variant of Qt), so all my code is in python (breadth and depth of backend ecosystem is unmatched, and code assistant AI knows python better than some alternatives across the breadth of needed experiences). Also pyside 6 keeps the option to rebuild for windows and linux apps easier.

This is my current thinking, but given the rapidly shifting landscape, might change as I find better alternatives (or better options get released).

Comments URL: https://news.ycombinator.com/item?id=46148656

Points: 1

# Comments: 0

Use code SHOWHN100 to download the installer for free (free for next few days) from the link. Drag the app into Applications folder and run it.

WHAT IS AHAI:

ahai is a 100% local private Mac app to find ideas scattered across markdown files (for me it was code repo READMEs, Obsidian notes, clipped web articles and research paper abstracts in Obsidian).

TECH DETAILS:

- GUI - Pyside 6 (Qt for Python)

- AI in app - mlx_lm

- backend - Python

- AI coding assistant - Claude Opus 4.5, Claude Code, Claude Code on web

- AI assistants for content - Grok 4.1 Thinking, Gemini 3 Pro, Nano Banana 2 pro

- System requirements - Mac with Apple Silicon, minimum 16GB unified RAM

BACKSTORY:

I have been researching how to find the balance between quality and acceleration while building using AI. Full on vibe coding is not for me. And going full manual doesn't make sense either. I finally found a formula that worked, and wanted to find an idea to work on end to end. I had lots of prototypes in my git repos, with READMEs describing the project ideas. I had jotted down ideas and clipped research papers to Obsidian notes (also markdown files). Totally over 13k markdown files - it was impossible to find the markdown files containing ideas, and ideas within them with any heuristic. I needed AI. So I wrote a script to do this using mlx models on Mac. It did so well, I decided to make that my first product. That is how ahai was born.

HOW IT WORKS:

- you point ahai to some folder, and it starts finding markdown files, and then uses AI (an mlx_lm model) to find if it has ideas, and then to extract ideas with title and a brief description.

- Clicking on an idea takes you to the rendered markdown source of the idea.

- You can then reorder the ideas, hide some of them, etc. and export the list of ideas to markdown, html or json.

- You can only be running one folder at a time. You can pause and resume folders.

- First time model use, if model is not already on your machine, will take some time to download. Be patient.

- You can change the model in settings if you know how to do that - must be an mlx_lm compatible model to work.

- All files are output to an output folder that you can also configure in settings. Switching between output folders can enable managing different kinds of stuff in different places - if you already downloaded some content in a folder, switch out and back, it will take off where you left off.

- Known issue: The ideas have false positives and false negatives. This is AI generated, cannot be avoided, but can be improved with prompting. Even with some of these, I find it quite useful.

- Known issue: Processing folders will take time, which is tuned to some degree, but cannot be avoided. But as I said, you can always pause and resume.

HOW IT IS DIFFERENT:

- Most AI apps and buzz focus on complex problems that only the best frontier models can solve, if any. I am interested in what kind of useful problems small local models can solve reliably. This app solves a niche problem using smaller local models very well. Most upcoming apps will also have the same focus.

- A lot of work has gone into benchmarking different models on markdown files to see which ones work best for a given size of machine (the app requires minimum 16GB RAM, but depending on the machine, it will decide which model to use as default). A tech/power user can always change the model used in the settings - just has to be an mlx_lm compatible model that fits in their RAM (within about half the size of total RAM).

- I have been using AI for coding and research and evals and all that, but until recently, it became hard to get anything work end to end as an indie dev - from concept to dev to marketing. But recently, with Claude Code/Claude Code web/Claude Opus 4.5, as well as Gemini 3 Pro/Nano banana 2 pro/NotebookLM deep research - I was able to build this app - with diligence in high risk parts, more trustingly in low stakes pieces - verifying everything, questioning anything suspicious - from concept to launch in 10 days.

- I think local private experiences are going to become increasingly relevant, as proprietary models and AI based apps suck in our data and can misuse/abuse/expose it in many ways. So, I believe this is a good space to focus on - local private Mac apps using local models. This is the first app in that space.

PRICING:

It is free with the code SHOWHN100 for this community for now, will be revoked at some point. Regularly priced at $19+, and suggested $29 - one time fee, no subscription, get all updates from later. I asked a bunch of top models by describing my app and they came up with this ballpark. I personally felt this was too pricey, but they also said a lower price would indicate poor quality to the users. Am open to changing it if there is evidence this isn't the right price point.

It is still rough on the edges. Please let me know any issues and I will prioritize and fix them.

Please try it out and let me know any questions. AMA on my tech stack, process, anything.

Comments URL: https://news.ycombinator.com/item?id=46120991

Points: 1

# Comments: 1