Dev Digest — April 25, 2026

🔥 HOT RELEASES

DeepSeek V4 — Open-Source Giant Drops DeepSeek V4 launched this week with 1.6T parameters, 1M context window, and open weights. V4-Pro rivals GPT-5.4 and Opus 4.6 on coding benchmarks (93.5 on LiveCodeBench, 3206 on Codeforces). V4-Flash is the cheap/fast variant for agents. API costs 10-50x less than OpenAI. The open-source AI arms race is accelerating hard. 🔗 X: https://x.com/polydao/status/2047525515880276324

GPT-5.5 Ships OpenAI's latest — stronger coding and agentic workflows, same speed as 5.4. Lives in both ChatGPT and Codex. Early takes: devs are calling it the most trustworthy model for engineering tasks. McKay Wrigley went from 80/20 Claude/GPT to 80/20 GPT/Claude in 3 months. 🔗 X: https://x.com/mckaywrigley/status/2047806060061507971

Claude Opus 4.7 — Mixed Reception Launched Apr 16 for complex reasoning and long-running agents, but community feedback is rough. Multiple prominent devs report degradation vs 4.6. Anthropic published a post-mortem acknowledging issues. Meanwhile, Claude Code 2.1.120 shipped with ultrareview command and 57% prompt token reduction. 🔗 X: https://x.com/omarsar0/status/2047378547518373924 🔗 X: https://x.com/ClaudeCodeLog/status/2047882241460613492

Hermes Agent v0.11.0 — "The Interface Release" Nous Research's big update: full terminal interface rebuild, unlimited recursive sub-agents, GPT-5.5 support, mid-task redirecting. Open source and gaining serious traction. 🔗 X: https://x.com/NousResearch/status/2047495677651918885 💻 GitHub: https://github.com/NousResearch/hermes-agent

Kimi Code (Moonshot AI) Drop-in Claude Code replacement using Kimi K2.6. Two env vars, 100 tokens/sec, 262K context window, included in standard Kimi membership ($0 extra). Personal use only for now. 🔗 X: https://x.com/0xRicker/status/2047992440750022836

Node.js v26 — V8 14.6 with Unflagged Temporal V8 14.6 landed in Node upstream, heading for v26. First version with unflagged Stage 4 Temporal API — the long-awaited replacement for JavaScript's terrible Date handling. 🔗 X: https://x.com/robpalmer2/status/2047746052653002999 💻 GitHub: https://github.com/nodejs/node/pull/61898

🧪 INTERESTING REPOS

MUX0 — Terminal Multiplexer for AI Coding Agents Sidebar management, tab management, split panes, with live status display for Claude Code, OpenCode, and Codex. Built on Ghostty renderer with full theme support. 🔗 X: https://x.com/Justin1024go/status/2047134550271410511 💻 GitHub: https://github.com/10xChengTu/MUX0 🌐 Site: https://mux0.com

ml-intern (Hugging Face) — Autonomous ML Research Agent An agent that reads papers, discovers datasets, writes training code, trains models, debugs failures, iterates, and ships to HF Hub. Passed HF's own internship test in 15 minutes. 5.8K+ stars. 🔗 X: https://x.com/akseljoonas/status/2047332440025321796 💻 GitHub: https://github.com/huggingface/ml-intern

gnhf ("Good Night, Have Fun") — Overnight AI Agent Orchestrator Fire a prompt, sleep, wake up to committed code. Auto git commits on success, auto rollback on failure, exponential backoff, worktree mode for parallel agents. Works with Claude Code and Codex. 🔗 X: https://x.com/GitHub_Daily/status/2047941193708040351 💻 GitHub: https://github.com/kunchenguid/gnhf

Agentlytics — Unified Analytics for AI Coding Tools Dashboard tracking usage across Cursor, Windsurf, and Claude Code. Finally see where your tokens are going. 🔗 X: https://x.com/tom_doerr/status/2047079888683683913 💻 GitHub: https://github.com/f/agentlytics

Kanna — Web UI for Claude Code & Codex CLIs Browser-based interface wrapping terminal coding agents. 🔗 X: https://x.com/tom_doerr/status/2048010768705093668 💻 GitHub: https://github.com/jakemor/kanna

OpenClicky — Open-Source Agent for Cursor MIT-licensed rebuild of the closed-source Clicky app. First push, actively developing. 🔗 X: https://x.com/jasonkneen/status/2047448944511230356 💻 GitHub: https://github.com/jasonkneen/openclicky

Andrej Karpathy Skills for Claude Code One CLAUDE.md file that "fixes Claude Code's worst habits." 78K stars. 🔗 X: https://x.com/RoundtableSpace/status/2047982700988891629 💻 GitHub: https://github.com/forrestchang/andrej-karpathy-skills

OpenRouter create-agent-tui A skill that builds your own agent harness + terminal UI. 4 customization paths, dozens of optional features. Route tasks across 300+ models. 🔗 X: https://x.com/OpenRouter/status/2047701992798392484

Token Cost Optimization Repos (Curated) 10 repos that cut Claude Code token costs by up to 97%: RTK (shell output compression), Context Mode (SQLite offloading), code-review-graph (Tree-sitter knowledge graph), Token Savior, Caveman (linguistic compression), and more. 🔗 X: https://x.com/Shruti_0810/status/2046621478020624589 💻 GitHub: https://github.com/rtk-ai/rtk 💻 GitHub: https://github.com/JuliusBrussee/caveman 💻 GitHub: https://github.com/ooples/token-optimizer-mcp

Awesome Agentic Patterns Curated catalogue of agentic AI design patterns — useful reference for anyone building agent systems. 🔗 X: https://x.com/tom_doerr/status/2047148676913504368 💻 GitHub: https://github.com/nibzard/awesome-agentic-patterns

💡 TECHNIQUES & IDEAS

Anthropic's MCP Production Guide New blog post: "Building agents that reach production systems with MCP." When to use APIs vs CLIs vs MCP, patterns for building servers, context-efficient clients, pairing MCP with skills. 🔗 X: https://x.com/ClaudeDevs/status/2047086372666921217 🌐 Blog: https://claude.com/blog/building-agents-that-reach-production-systems-with-mcp

12 MCP Patterns Behind Production Agents @bibryam's deep-dive: intent-based tools (not endpoint mirrors), remote-first servers, inline UI, discoverable OAuth, context economy, plugin + skill distribution. 🔗 X: https://x.com/bibryam/status/2048044127690846365 🌐 Blog: https://generativeprogrammer.com/p/12-mcp-patterns-behind-production

Dual-Agent Workflow: Claude Code + Codex Side by Side Practical guide: port CLAUDE.md to AGENTS.md, wire MCPs to both, route tasks by strength. Opus for deep reasoning/long context, GPT-5.5 for speed and simple refactors. 🔗 X: https://x.com/Axel_bitblaze69/status/2048004308793524392

Obsidian × Claude Code Knowledge Base 3-layer architecture (raw → wiki → schema) with 4 operation cycles (Ingest, Compile, Query, Lint). Building a personal Wikipedia with AI as editor. 🔗 X: https://x.com/yusaku_0426/status/2047910777261350970

Rubydex MCP — Does It Actually Save Tokens? Real-world test on Timee's monolith Rails app measuring whether Rubydex MCP reduces Claude Code token consumption. 🔗 X: https://x.com/euglena1215/status/2047978755906474437 🌐 Blog: https://tech.timee.co.jp/entry/2026/04/25/185149

🔮 EMERGING TRENDS

The Great Coding Agent Fragmentation The era of one dominant coding agent is over. Claude Code, Codex, Hermes Agent, OpenCode, Kimi Code, Cursor — devs are running multiple agents side by side, routing tasks by model strength. The meta is becoming model-agnostic orchestration.

Open-Source Models Reaching Parity — Weekly Three frontier-class open models shipped in one week: Kimi K2.6, Qwen 3.6, DeepSeek V4. All free, all runnable locally. The "competitive moat of closed models" argument is getting harder to make each month.

Token Optimization as a Category A whole ecosystem of tools is emerging just to reduce AI coding costs — shell output compressors, context offloaders, linguistic compressors, codebase knowledge graphs. Token cost management is becoming its own discipline.

Anthropic's MCP Becoming Infrastructure Standard MCP adoption is accelerating across all major coding agents. Anthropic's new production patterns blog signals it's maturing from "cool protocol" to "real infrastructure." Agent skills and tool distribution are the next layer.

The "AI Intern" Archetype HuggingFace's ml-intern represents a new pattern: autonomous agents that complete entire research workflows end-to-end (read papers → train models → ship results). Expect more domain-specific "intern" agents.

April has been the most intense month for AI model releases in history. In 24 days: DeepSeek V4, GPT-5.5, Opus 4.7, Qwen 3.6, Kimi K2.6, Llama 4, GLM-5.1, Claude Mythos Preview, Gemma 4. One week left.

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