Local/offline & desktop-agent surge

Massive ecosystem growth in local/offline AI agents. Key new: Kimi K3 open-weights model (Moonshot AI) — frontier-level, matches Claude Fable and GPT-5.6 Sol, free to download/deploy but requires 1.5TB VRAM or 3TB RAM, impractical for personal use. Muse Spark 1.1 (Meta AI) — free Thinking mode, strong computer use, game creation, now SOTA on video-to-code. Second Brain for AI v2 — free open-source memory layer across agents. FluencyLoop — free open-source tool teaching at your skill level, auditing changes. Postfleet — free tier AI agent email security with MCP. Also: Codux, Gepard 1.0, Kokoro, Convos, Manta AI, Tabula, Buselligence, Superpowers, and many more. New: NVIDIA fully open-sourced Cosmos 3 Edge (weights, recipes, code) — free for side projects. Thinking Machines open-sourced Inkling, a native multimodal model. New: ProtoFlow — free AI-powered PCB schematic design tool. New: MyPeople — free self-hosted orchestration board for managing multiple AI coding agents (Claude Code, Codex, Grok) with a persistent Boss agent; launches today. New: Laguna S 2.1 (poolsideai) — 118B MoE with 8B active, 1M context, thinking mode, runs FP8 on 128GB machines (great for Macs), free to run if you have hardware. Soumith Chintala endorses it for agentic use, fits on DGX Spark. New articles: Agno vs AutoGen comparison (free frameworks), Claude Code remote Mac setup guide, 'How to Build Your First Agent' guide, 'Learn AI Agents in 2026' course list. New: Stack Exchange Q&A on running a local AI personal assistant (Ollama, agents, function calling) — practical guide for side projects. New: OpenCode Superapp — free tool bringing Codex-like agentic power to local/self-hosted models with voice and computer use. New: Plasma AI Fractal — open-source recursive agent architecture for large-scale coding (Apache 2.0). New: AgentLoop — open-source coding agent loop with fresh-context-per-cycle, avoids context decay. New: Gemma 4 updated with better tool calling and reasoning — tested locally with llama.cpp, free to download (26B weights). New: Andrew Ng's OpenWorker — open-source, local-first desktop agent that returns finished deliverables instead of chat (MIT license, BYO model). New: J.A.R.V.I.S — free Windows AI assistant bundling multiple models with RAG, voice control, automation, local-first. New: Screenpipe — open-source local screen/audio recorder that gives AI agents persistent memory; privacy-focused but has performance overhead on older Macs. New: Cost-saving tip: switching models mid-session resets prompt cache, inflating bills. Open-weight router from @daridotdev only switches when it saves money, plugging into Claude Code/Codex. New: Engram Vault Sync — free Obsidian plugin that turns notes into AI-accessible memory via MCP, self-hostable, no telemetry. New: CodexBar Lite — free macOS menu bar tool for tracking Codex usage, privacy-first. New: Practical comparison of three local coding models (GPT-OSS, etc.) — GPT-OSS won despite being smallest, challenging bigger-is-better assumption. New: OpenHuman — comprehensive open-source personal AI assistant, local-first, with persistent memory, orchestration, and deep research. Memory tree and subconscious background loop are standout features. Free and privacy-focused. New: OpenAgent — open-source personal AI agent platform connecting to every major LLM, supports browser use, RAG, workflow automation, and MCP. Free to self-host, covers many personal productivity and side project use cases. New: herdr — free open-source terminal multiplexer for managing multiple AI coding agents (Claude Code, Codex, etc.), 22.5k stars, Rust binary. Practical for side project developers. This surge continues to provide free, self-hosted options for side projects.