Coding agents

• mksglu/context-mode sandboxes MCP tool output (claimed 98% reduction, e.g. 56 KB Playwright snapshots) and persists session memory across compaction, via MCP + hooks on 17 platforms.
• colbymchenry/codegraph is a local, auto-syncing code knowledge graph for Claude Code, Codex, Cursor, Gemini and others, aimed at fewer tokens and fewer tool calls.
• DietrichGebert/ponytail is a skill that makes the agent write less code; claims ~54% less code and ~20% cheaper on Haiku 4.5 (n=4, 12 tasks), so treat the numbers as thin.
• HydraFusion in VS Code and the GitHub Copilot app moves the research preview beyond Copilot CLI and into the model picker.
• modelcontextprotocol/servers is the reference MCP server repo; community servers now live in the MCP Registry instead.
• openclaw/openclaw is an open-source personal assistant that runs locally and reaches you via Discord, iMessage, Slack, Telegram, WhatsApp and 20+ other channels.
• ComposioHQ/awesome-claude-skills is a 1000+ item skills/plugins list spanning Claude Code, Codex, Cursor and Gemini CLI, pushing Composio's MCP Gateway.
• Super excited to be interviewing @poteto announces a live Friday 9AM PT talk on skills, software factories and shipping with agents.

Models and labs

• Introducing Gemini 4 Argon is a new frontier model for coding, enterprise work and cyber defense, limited to trusted testers via the Fairwind Program.
• RT @NerdSnipePod: Opus 5.5 is a long podcast episode covering Opus 5.5, GPT-6 Sol & Luna, Grok 4.7, why model routing fails, and pricing and usage limits.
• Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week pushes back on Dwarkesh's RLVR framing of computer use and covers OpenAI's quick Jev response.
• Disrupting a coordinated model-distillation campaign describes OpenAI shutting down an effort to extract protected model reasoning and hardening against adversarial distillation.
• https://t.co/sRyO7OigQ7 is our new home for developers building with Claude is a new Anthropic developer hub with engineering deep dives, Claude Code and API guides.
• RT @ClaudeDevs: new home for developers building with Claude is a retweet of the same Anthropic developer hub announcement.

Local inference and CLI/TUI

• Launch HN: Magnitude (YC S25) is an inference engine that tunes its kernels to your hardware and claims up to 2x faster than llama.cpp; HN skeptics note MLX already exists and no methodology was given.
• T3 Code now supports Sign in with ChatGPT, including multiple accounts and no external dependencies.

Vibe-coding

• Not a full-time dev: shipped a 3-language Android app in 3 weeks is a solo Claude Code + Claude Design field report on what worked and what bit back.
• How many non-developer vibecoders finally have configs sophisticated enough to one-shot anything is a Reddit thread sampling how far non-developers' agent configs now go.
• jambuild offers near-instant multiplayer vibecoding by pointing and talking.

Agentic SaaS

• heygen-com/hyperframes renders deterministic MP4 from HTML/CSS, driven by agents, with a Claude Code plugin.
• Campfire is a shared workspace for humans and coding agents.
• GitBot lets you build bots on top of the coding agent you already use.
• m'kay is a single voice interface from your phone to all your coding agents.
• CrawlRaven MCP exposes SEO work to your AI agent over MCP.
• Jakeschincariol/linkedin-agent-skill is eleven free Claude skills for running a LinkedIn account, including a humanizer that scores drafts.