Dev Digest — March 26, 2026

🔥 HOT RELEASES

Voxtral TTS by Mistral AI Open-weight text-to-speech model that reportedly outperforms ElevenLabs v2.5 Flash. ~70ms latency, 9 languages, voice cloning from 10 seconds of audio, emotionally expressive speech. 3B params, open weights. 🔗 X: https://x.com/MistralAI/status/2037183026539483288

Cohere Transcribe New open-source speech-to-text model — #1 on HuggingFace Open ASR leaderboard with 5.42% word error rate. 2B params, Apache 2.0, 14 languages. Day-0 vLLM support with 2x throughput optimizations for encoder-decoder models. 🔗 X: https://x.com/cohere/status/2037159129345614174 💻 HuggingFace: https://huggingface.co/blog/CohereLabs/cohere-transcribe-03-2026-release

Cline Kanban — Multi-Agent Orchestration Cline released a standalone app for running multiple CLI coding agents in parallel. Each task gets its own git worktree, live diffs for review, dependency chains between cards. Works with Claude Code, Codex, and Cline CLI. Free and open-source. 🔗 X: https://x.com/cline/status/2037182739695493399

Claude Code Auto-Fix Claude Code now has an "Auto Fix" setting — it automatically fixes your PR in the background. Turn it on and walk away. 🔗 X: https://x.com/lydiahallie/status/2037238977133240713

Vibe Coding XR by Google Research New prototyping workflow combining Gemini Canvas with XR Blocks framework — turn prompts into interactive, physics-aware WebXR applications. 🔗 X: https://x.com/GoogleResearch/status/2036904458974871944

🧪 INTERESTING REPOS

JSSE — JavaScript Engine Built Entirely by Claude A from-scratch JS engine in Rust, built by Claude Code in YOLO mode over 6 weeks with only 4 hours of human supervision. Passes 100% of test262 (98,426 scenarios). ~$4k in Opus 4.6 API calls. Zero lines of Rust written by a human. 🔗 X: https://x.com/robpalmer2/status/2037093737088618637 💻 GitHub: https://github.com/pmatos/jsse

agenticSeek — Fully Local Manus Alternative Autonomous web browsing, code execution (Python, Go, C, Java), voice input, multi-agent task planning — 100% on your hardware. No API keys, no subscriptions. 🔗 X: https://x.com/_vmlops/status/2037136534747701765 💻 GitHub: https://github.com/Fosowl/agenticSeek

LiteParse — Model-Free Document Parser Open-source parser by LlamaIndex that digitizes text from any PDF in seconds. One-line installable skill for Claude Code and 40+ other agents. Great for coding agents that can't read PDFs natively. 🔗 X: https://x.com/jerryjliu0/status/2036960731871277353 💻 GitHub: https://github.com/run-llama/liteparse

AI Website Cloner for Claude Code Claude Code skill that clones entire websites from one prompt — uses Chrome MCP to pull assets/code directly, spawns parallel agents for sections. 🔗 X: https://x.com/RoundtableSpace/status/2037095965308367282 💻 GitHub: https://github.com/JCodesMore/ai-website-cloner-template

Telegram Drive Open-source desktop app (Tauri + Rust + React) that turns your Telegram account into unlimited cloud storage. 🔗 X: https://x.com/akaclandestine/status/2037189403693531297 💻 GitHub: https://github.com/caamer20/Telegram-Drive

TaxHacker — Self-Hosted AI Accounting Upload receipts/invoices in any language → AI extracts everything → auto-categorizes into tax buckets. Custom LLM prompts, multi-currency (including crypto), CSV export. MIT License. 🔗 X: https://x.com/heynavtoor/status/2035940533353562351 💻 GitHub: https://github.com/vas3k/TaxHacker

Awesome Agents Comprehensive list of 30+ agent frameworks (CrewAI, AutoGen, Swarm, Mastra), coding agents (Cline, OpenHands, Aider), research/browser/voice tools. 🔗 X: https://x.com/_vmlops/status/2037110737806782798 💻 GitHub: https://github.com/kyrolabs/awesome-agents

Claude Prism — Offline Scientific Writing Write scientific papers with Claude AI, fully offline workflow. 🔗 X: https://x.com/tom_doerr/status/2037128951009550346 💻 GitHub: https://github.com/delibae/claude-prism

Understand Anything Turns complex concepts into first-principles explanations. Structured clarity for code, AI, math — anything. 🔗 X: https://x.com/RoundtableSpace/status/2037231861572002297 💻 GitHub: https://github.com/Lum1104/Understand-Anything

🎥 WORTH WATCHING

Browser Use: Best AI Researcher Benchmark Browser Use claims they beat OpenAI, Google, and Anthropic on the biggest browser agent benchmark. Their researcher is an LLM + CLI that spots its own failures, finds root causes, and self-corrects in a loop. Video demo in the tweet. 🔗 X: https://x.com/browser_use/status/2037190339975397485

freeCodeCamp: Stop Letting AI Agents Guess Your Requirements Tutorial on spec-writer, a Claude Code skill that turns rough feature requests into specs + technical plans + task breakdowns before your agent starts coding. 🔗 X: https://x.com/freeCodeCamp/status/2037138104516006313 🔗 Article: https://www.freecodecamp.org/news/how-to-stop-letting-ai-agents-guess-your-requirements/

Codex App Server — 100% Open Source Reminder that OpenAI's Codex App Server is fully open source. Build on top of Codex with ChatGPT Auth. Includes VS Code extension. 🔗 X: https://x.com/reach_vb/status/2036904822641676716

💡 TECHNIQUES & IDEAS

dmux — Never Lose Your Coding Agent Again Wrapper around tmux that keeps your coding agent running even when you close your terminal. Just type dmux to reconnect exactly where you left off. 🔗 X: https://x.com/jpschroeder/status/2037188194257580061 🔗 Site: https://dmux.ai

Kilo AI Leaderboard — Nuanced Model Rankings Founded by ex-GitLab CEO, Kilo breaks down model performance by coding task type (Architect, Debug, Code). Xiaomi's MimoV2 leads Architect, MiniMax leads Debug. More targeted than OpenRouter for dev use cases. 🔗 X: https://x.com/kevinsxu/status/2037186092802130060 🔗 Leaderboard: https://kilo.ai/leaderboard

DefiLlama MCP 23 tools bringing on-chain DeFi data to your AI agent with guided research workflows for structured analysis. 🔗 X: https://x.com/DefiLlama/status/2037234604629753954

🔮 EMERGING TRENDS

Speech AI Going Open — Both Mistral (TTS) and Cohere (STT) dropped open-weight/open-source speech models on the same day. The full voice stack is rapidly commoditizing. Mistral's Voxtral reportedly beats ElevenLabs; Cohere's Transcribe beats Whisper.

Multi-Agent Orchestration Maturing — Cline's Kanban brings proper project management to AI coding agents. Git worktrees for isolation, dependency chains, live diffs. The "run 10 terminals" era is ending.

AI-Built Software Getting Serious — JSSE (a full JS engine passing 100% of test262) was built entirely by Claude with minimal human oversight. We're past toy demos — agents are shipping production-grade systems.

GitHub Under Pressure — Wes Bos noted GitHub is struggling under the weight of every AI platform building on top of their permissive GraphQL API. The infrastructure layer is feeling the AI agent load.

Full digest: https://digest.99.cooking