GPT–LLM native macOS app with time travel, versioning, search
Hey everyone! I made a Mac app for exploring large language models. It’s fast, small, has a tiny memory footprint. It’s immutable by design with both immediate time travel and automatic versioning as foundational elements. The app is written in Swift and a bit of Rust for the tokenizer. I used SwiftUI for structure and animations and Cocoa for advanced behavior. All storage is SQLite and local-only. You can go through the database as needed and backup it as well. The app has support for variants, which is the `n` parameter in the OpenAI chat completion API—equivalent to the drafts feature in…
In plain words
This macOS app lets users explore and experiment with large language models through an interface built in Swift and Rust. Designed for efficiency with a small memory footprint, it stores all data locally in SQLite and features time travel, automatic versioning, and variant generation to compare different model outputs. The app displays token counts and API costs, supports marking and hiding runs, and allows users to browse the history of conversations. It's built for anyone working with LLMs who wants fast, lightweight exploration with full control over their data.
written from the facts on this page · September 2026
From the sources
In the maker’s words, at launch
Hey everyone! I made a Mac app for exploring large language models. It’s fast, small, has a tiny memory footprint. It’s immutable by design with both immediate time travel and automatic versioning as foundational elements. The app is written in Swift and a bit of Rust for the tokenizer. I used SwiftUI for structure and animations and Cocoa for advanced behavior. All storage is SQLite and local-only. You can go through the database as needed and backup it as well. The app has support for variants, which is the `n` parameter in the OpenAI chat completion API—equivalent to the drafts feature in Bard. It shows model- and ChatML-aware tokens and cost, and it dynamically adjusts response length in function of prompt length and context length. It supports hiding runs that are not useful, marking the ones that are notable, and going through the ancestry of the current run. It also has support for examples, as specified in ChatML, conversation names and personal and local-only notes. The full-text search works across all stored text and it has support for `all`, `any`, `prefix` and `phrase` matching; the results are time-ordered rather than ranked for the time being. There are several shortcuts and more are coming so that the app can be used entirely by keyboard. You can export to JSON and all data is exported. The app is sandboxed and notarized. The API keys are stored securely in the Keychain. Next I plan to implement combinatorial runs (mixing multiple values of the same parameter and multiple models as well) and full Markdown support. Currently only OpenAI’s conversational models are supported (GPT 3.5, GPT 4 8K, GPT 4 32K), but I’ll be adding support for local models and custom ChatML endpoints as soon as possible. Give it a try. Any and all feedback welcome! Thoughts and questions as well! William
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 27d ago · cactuscompute.com
Launched alongside, April 2023
the whole month →
- G4
Hi HN, Today we’re launching GPT-4 answers on Phind.com, a developer-focused search engine that uses generative AI to browse the web and answer technical questions, complete with code examples and detailed explanations. Unlike vanilla GPT-4, Phind feeds in relevant websites and technical documentation, reducing the model’s hallucination and keeping it up-to-date. To use it, simply enable the “Expert” toggle before doing a search. GPT-4 is making a night-and-day difference in terms of answer quality. For a question like “How can I RLHF a LLaMa model”, Phind in Expert mode delivers a…
AI · 2023 · phind.com
