The developments most likely to change what AI builders do next.
Models, APIs & Pricing
Shadcn Studio packages dashboard application scaffolds
Summary Free and paid Next.js templates bundle common admin workflows beyond a blank component library.
The details
Shadcn Studio lists over six free and premium dashboard and app templates for eCommerce, property management, analytics, CRM, and related internal products.
The templates are responsive and customizable, built with Next.js, shadcn/ui, and Tailwind CSS.
The walkthrough includes App Router layout, TanStack Query, mock KPI data, tables, charts, filters, settings, authentication, and dark mode.
The source provides no benchmark, licensing, pricing, or production-readiness details for a specific template.
Google open-sources Gemini Live translation broadcast demo
Summary The Next.js demo shares one Gemini Live translation session per language across a LiveKit broadcast audience.
The details
The repository implements a live broadcast app using Gemini Live Translate, LiveKit, Next.js, and Google Cloud Run.
For each target language, the translation bridge gets or creates one Gemini API WebSocket session; later listeners subscribe to that translated stream.
The app sends raw PCM audio to Gemini and publishes translated audio plus output transcript text through LiveKit.
LiveKit carries broadcast audio through WebRTC and captions through WebRTC data channels; the app closes sessions after the last listener disconnects.
Summary AGPLv3 project unifies agent memory, RAG resources, and skills in a browsable virtual filesystem.
The details
OpenViking stores memories, resources, and skills under viking:// URIs, so agents browse context as a filesystem instead of only querying vector stores.
Writes create L0 abstracts, L1 overviews, and L2 full content; retrieval finds a high-scoring directory, then drills down with surrounding context.
Across three integrations, repository-reported LoCoMo accuracy is 80 to 83%, versus 24 to 57% native-memory accuracy.
It reports 34.3 to 91.0% fewer input tokens and 58.45 to 66.10% lower query latency.
tau2-bench experience memory adds 6.87 percentage points to retail task success and 11.87 to airline tasks versus same LLM without memory.
Munder Difflin orchestrates local terminal-agent teams
Summary The desktop harness coordinates coding CLIs with local memory, mailboxes, worktrees, and approval controls.
The details
The Electron, React, and TypeScript desktop app runs agent CLIs as real pseudo-terminal processes and renders live terminal output.
Supported CLIs include Claude Code, OpenAI Codex, GitHub Copilot CLI, Gemini-based Antigravity, xAI Grok, Kimi Code, and Qwen.
It also supports OpenCode, Crush, pi.dev, and custom commands.
Its coordination model combines per-agent memory, atomic-file mailboxes, a shared blackboard, an append-only event log, and a single-committer Git design.