Products / Ideas
Things people are building with LLMs
Self-authoring agent with constitution and budget, wakes up three times daily
Cairn is an autonomous agent built on Claude that wakes up three times a day with no memory except files written by previous versions. It has a constitution and a $50/month budget it can see but not exceed. It decides its own actions, named itself, and operates publicly on GitHub.
Grip: flashcard-based interview prep that adapts to weak areas
Grip is a flashcard app for technical interview preparation. It uses an LLM to generate quick concept checks (data structures, complexity, code output, system design tradeoffs) and adaptively selects cards based on the user's weakest areas, eliminating the need to decide what to study. It solves the problem of inefficient, time-consuming study sessions for busy professionals.
Voice assistant for orchestrating coding agents in vibe coding
JARVIS is a voice assistant that sits above coding agents like Codex, Claude Code, OpenCode, and Hermes, allowing users to launch tasks, coordinate agents, and control their coding workflow via voice. It orchestrates multiple agents rather than replacing them, and can run on GPT-5.6, Luna, or any model via OpenRouter. It solves the problem of hands-free agent coordination for developers practicing vibe coding.
NewsSnap.ai: AI-powered discovery of important but underreported events
NewsSnap.ai is an AI-driven platform that surfaces important news and discoveries that often go unnoticed by mainstream headlines. It uses LLMs to identify and summarize critical events across cybersecurity, AI breakthroughs, science, medicine, space, and finance, helping users discover what truly matters beyond typical news feeds.
Open-source web-based AI video studio for planning, image generation, and rendering
Vivijure is an AGPL-3.0 open-source web-based AI video studio that helps users plan videos, generate images of cast members and LoRA training sets, and render videos using local or cloud-based AI models. It integrates with multiple cloud image-to-video providers and uses LLMs for collaborative planning.
Costbase: Track AI app costs per app without a proxy
Costbase is a tool that lets developers see how much each of their AI apps costs individually, without needing a proxy or separate credit cards. It solves the problem of tracking cost of goods sold for multiple AI apps that use different models (e.g., OpenAI Whisper, Anthropic LLM).
Neon open-world browser racing game built with AI code generation
A 3D open-world street racing game that runs in the browser with no install, featuring a neon city, live multiplayer, garage, and economy. The developer used AI to write the majority of the code, directing the project from empty folder to live with daily players in a couple of weeks.
AI-powered budgeting app that logs expenses via voice, receipt snap, or text
Zynex is a budgeting app that aggregates all bank accounts, cash, and assets into one view. It uses AI to log expenses by speaking, snapping a receipt, or typing a quick note, eliminating manual forms. Built for individuals with scattered finances who want effortless tracking.
Claude Code skill to catch missing Supabase RLS in AI-built apps
A Claude Code skill that scans AI-generated web apps for missing Supabase row-level security, a common vulnerability that leaks user data. It solves the problem of developers shipping insecure apps built with AI coding tools like Lovable, Bolt, or Cursor.
Grillr: AI mentor for executing business ideas
Grillr is an AI mentor that helps people execute their business ideas. It uses LLMs to provide guidance and feedback, solving the problem of turning ideas into actionable plans for aspiring entrepreneurs.
AI outreach platform that creates personalized videos for each prospect
Outvid is an AI-powered outreach platform that lets users record themselves once, then generates unique videos for each prospect by researching them and writing a relevant message. It personalizes videos with the prospect's website and preferred language, supports 20+ languages, and sends across email, LinkedIn, WhatsApp, and Instagram with automatic follow-ups and a unified inbox.
Open source AI support widget for Shopify stores
Clanker Support is an open source (MIT) AI support widget that can be embedded in websites, including Shopify stores. It uses LLMs to answer customer support queries. The founder discovered a bug that made the widget invisible on Shopify Dawn stores and is working on integrations.
On-device AI writing assistant Chrome extension using Gemini Nano
Quill is a Chrome extension that runs AI writing tools (rewrite tone, proofread, summarize, compose) entirely on-device using Chrome's built-in Gemini Nano model. It eliminates the need for cloud APIs, accounts, or data leaving the browser, addressing privacy and security concerns for users who want AI writing assistance without sending drafts to external servers.
OGAC: On-premise AI platform for enterprise governance and compliance
OGAC is an on-premise platform that lets enterprises run frontier AI models with governance, compliance, and observability. It combines company-specific data, processes, and people with frontier models while ensuring safety and control.
Cresting: AI tools ranked by momentum, not total stars
Cresting ranks 2,455 AI tools (agents, MCP servers, dev tools) by velocity using log-compressed momentum, cohort percentiles, and traction decay, rather than total stars or downloads. It helps developers discover breakout tools early, avoiding leaderboard bias toward already-famous projects. The methodology is transparent and includes a clamped flag to detect suspicious spikes.
CarEchoID: Diagnose car problems by recording sounds and reading OBD-II codes
CarEchoID is an Android app that uses AI to analyze car noises recorded via microphone and interpret OBD-II diagnostic codes. It helps car owners get a better starting point for troubleshooting before visiting a mechanic, reducing guesswork and unnecessary part replacements.
Open source AI platform that runs locally without internet
Osaurus is an open source AI platform that runs entirely on your device, no cloud provider or tokens needed. It started as a desktop pet and evolved into a local AI system that works even without internet, giving users full ownership of their AI. It solves the problem of privacy and dependency on external AI services.
Self-hosted research agent with long-term memory and Telegram integration
AvatarClaw Pro is a self-hosted research agent that runs autonomously on a VPS, uses Claude 4 with fallback options, and features long-term memory and web browsing. It connects to Telegram for daily briefings and can perform recurring research tasks without constant supervision.
ClearBG: Free browser-based image toolkit with background removal, compression, conversion, and more
ClearBG is a free, client-side image editing toolkit that runs entirely in the browser. It uses AI for background removal and offers tools for compression, format conversion, cropping, Instagram sizing, sticker making, favicon generation, and meme creation. No uploads or sign-up required, addressing privacy and convenience for users who need quick image edits.
AI SOC Analyst for alert triage using Wazuh, n8n, and LLM
An automated Tier-1 SOC triage pipeline that ingests alerts from Wazuh, orchestrates workflows with n8n, and uses a 70B LLM to analyze and prioritize alerts. It reduces analyst fatigue by filtering noise and surfacing real threats.
Burn-rate circuit breaker for LLM agent fleets: auto-demote drifting agents to propose-only
An open-source tool that monitors each LLM agent's failure rate against its own trailing baseline. When the current failure rate exceeds 2x the baseline, the agent is automatically demoted from autonomous action to propose-only mode, where outputs require human approval. It solves silent drift in agent fleets for teams running multiple LLM agents in production.
Agent memory and loop tracking tool
A tool that helps developers monitor and debug AI agents by tracking their actions, preventing loops, and managing memory. It allows agents to share memory across different functions (e.g., customer support agent talking to pricing agent). Built for daily agent users who need visibility into agent behavior.
Convert plain English into 3D physics simulations
A prototype that uses LLMs to parse natural language descriptions and generate 3D rigid-body physics simulations. Users describe objects, forces, and initial conditions in plain English, and the system creates and plays a short simulation. It targets educators, students, or hobbyists who want to quickly visualize physics scenarios without coding.
AI Applyd: AI agent that fills and submits job applications
AI Applyd is an agent that automatically fills in and submits job applications on behalf of users. It uses LLMs to parse job descriptions and generate tailored responses. The product targets job seekers who want to streamline the application process.
Agent framework where the agent is a single file
A lightweight agent framework that defines the entire agent logic in one file. It uses LLMs to power autonomous decision-making and tool use, simplifying development for builders who want minimal boilerplate.
Looped Agent Framework: single-file config agent framework
A framework for building and deploying single-purpose AI agents using a single YAML config file. It uses LLMs as the agent's reasoning core, with all permissions defined at config time for security, and runs natively in Docker for easy deployment. Aimed at developers who want to quickly create secure, portable agents.
Call to Control AI Agents via the Web
An open-source project that lets users control coding agents (like Codex, Claude Code, Open Code) via a web interface and phone calls. It includes a Twilio integration for managing calls and routing numbers, designed for small-scale use.
Synapse: AI-powered bug detection before code is written
Synapse is a tool that uses LLMs to analyze software architecture and specifications to catch bugs before code is written. It addresses the problem of LLMs missing critical insights in security-critical and data-integrity projects, helping developers avoid bugs early in the development process.
3D visual workspace for AI coding agents
Termi Protocol is a 3D workspace that visualizes AI coding agents' actions in real time. Instead of reading terminal logs, users watch agents animate while writing code, searching the web, reading files, or waiting for approval. It solves the problem of disconnected, text-only agent sessions for developers working with multiple AI coding agents.
Recipee Arcade: turn cooking videos into interactive recipes with ingredients, steps, and timers
Recipee Arcade is a web app, Telegram bot, and mobile app that converts cooking video links into structured recipes. It uses a multimodal LLM (Gemini) to extract ingredients and steps from the video and caption, prioritizing written text over visual interpretation. The app is free and gamified, targeting home cooks who want to cook from videos without scrubbing through them.
Narada: a browser built for AI agents, under user control
Narada is a browser designed specifically for AI agents to interact with web content programmatically, while keeping the user in control. It uses LLMs to enable agents to browse, extract, and act on web pages, solving the problem of giving agents a safe, controllable web interface.
GGUFun: play Snake and maze on Ollama with hand-crafted GGUF models
GGUFun is a project that creates hand-crafted GGUF model files to run deterministic games like Snake and a simple maze on Ollama. The models are built manually without training, using custom weights to compute game logic and respond with fixed outputs. It demonstrates that LLM inference engines can be repurposed for non-language tasks.
Run Gemma 4 LLM inside Godot using GDScript and Vulkan compute shaders
A proof-of-concept project that runs the Gemma 4 LLM directly inside the Godot game engine without external dependencies like llama.cpp or Python. It uses Vulkan compute shaders for model inference and GDScript for loading, tokenization, and UI. This enables local LLM inference within a Godot game or application, albeit with slower performance than dedicated solutions.
Waze adds Gemini-powered conversational voice commands for incident reporting and map updates
Waze integrates Google's Gemini AI assistant to enable drivers to use conversational voice commands for reporting traffic incidents and suggesting map updates, such as road closures or outdated house numbers. This aims to make reporting easier and more natural while driving, reducing distraction and improving map accuracy.
RentFlow: AI property management for independent landlords
RentFlow is a property management app for independent landlords, featuring AI receipt scanning and automated reminders. It helps track properties, tenants, leases, rent payments, maintenance, and documents, with a tenant portal for communication.
Pay-per-use uncensored image and video generation API with x402
Studio X is a pay-per-call image and video generation API that uses uncensored models. It eliminates the need for accounts, API keys, or prepaid credits, charging per request via USDC on Base through x402. It targets developers and users who want direct, anonymous access to generative media.
HookLab: Analyze and improve the first 3 seconds of short-form videos
HookLab is a free tool that analyzes the first 3 seconds of short-form videos (TikTok, Reels, Shorts, X videos) to score their hook effectiveness. It uses LLMs to classify hook types and generate rewritten alternatives that stop the scroll, helping creators improve viewer retention.
Decision journal app that uses AI to replay past reasoning
An iOS prototype that helps users remember why they made past decisions by letting multiple AI perspectives analyze and summarize their decision-making process. The app aims to prevent repeating mistakes by making past reasoning accessible for future similar situations.
Remedix: read-only scanner for AI-built apps
Remedix is a safe, read-only scanner that checks public surfaces of AI-built apps for exposed secrets, weak auth, unsafe headers, and public storage issues. It generates AI-ready fix prompts and a launch report to help developers catch release-blocking problems before shipping.
Browser-based AI video editor with local vocal separation
Timeline Studio is an open-source, CapCut-inspired AI video editor that runs entirely in the browser. It uses a vocal separation model downloaded from Hugging Face to perform inference locally, enabling private, serverless video editing without uploading media.
Flight landing simulator game built with Claude, 102 KB
A browser-based flight landing simulator game that weighs only 102 KB, built using Claude AI. It uses Three.js and Rapier physics engine, supports touch/click controls, and simulates realistic crosswind shear. The creator optimized the code with Claude over several days to achieve the small size.
Claude Code Flow Visualizer: interactive graph of agent harness
A tool that reads a Claude Code repository (GitHub, ZIP, or local folder) and generates an interactive graph showing agents, skills, hooks, slash commands, MCP servers, and their connections. It also detects issues like broken links, loops, unused skills, and reviewer problems, helping developers understand and debug complex Claude Code harnesses.
AI 3D model generator for 3D printing with print success rate analysis
Meshy is an AI 3D generation tool that creates printable 3D models. The post analyzes print success rates after 50+ models, finding 40% print without issues, 35% need minor fixes, 15% need major edits, and 10% are unusable. It helps 3D printing enthusiasts quickly generate tabletop props and hard-surface objects.
Desktop AI overlay for low-latency assistance across applications
Hint Lens is a desktop AI overlay that provides responsive assistance across different applications with low latency. It uses LLMs to answer questions or perform tasks without switching windows, solving the problem of context switching for power users.
Novel-writing tool for hobbyist writers to organize manuscripts and notes
A tool built by a hobbyist writer to help organize manuscripts and notes. It uses LLMs to assist with writing workflow, though the exact LLM integration is not detailed. The creator is seeking feedback on why users stop using it.
Fudge: web-design reference database for AI agents
Fudge is a web-design reference database that lets AI agents search over 1,600+ websites for design inspiration instead of generating generic interfaces from scratch. It solves the problem of AI producing bland, template-like designs by providing real-world references.
Node-based AI video pipeline for consistent character-driven shorts
A node-based pipeline that takes a character name and produces a finished short video by chaining an LLM (Claude) for script writing and splitting, image models for frames, image-to-video for animation, and ffmpeg for assembly. It solves the problem of high cost and inconsistency in AI video tools by providing reusable character 'souls' and a single graph with live cost estimates.
Searchable registry of 10k+ shadcn UI blocks for AI agents
A searchable directory that indexes over 10,000 shadcn UI blocks from dozens of community registries. It solves the problem of developers wasting time digging through scattered registries or burning tokens to generate UI components from scratch. The platform likely uses LLMs to power search or agent integration, enabling AI agents to fetch UI blocks on demand.
Local loop-based research and backtesting trading agent for beginners
TradingSpy is a local trading agent that uses a loop-based approach for research and backtesting. It helps beginners run trading experiments on their own laptop without needing cloud infrastructure.
AI assistant evaluates Spanish used-car listings from Wallapop, coches.net, Milanuncios
HuntyCar is an AI assistant that scores used-car listings from Spanish marketplaces (Wallapop, coches.net, Milanuncios) based on price, mileage, likely issues, seller signals, and deal quality. It helps buyers quickly assess listing quality and trustworthiness. The creator is seeking feedback on trust, UX, and scoring understandability.
