AI Tools Daily

Claude Code 2.1.228 hardens synced Skills against supply-chain attacks (plus a Write-tool change worth knowing), auto mode becomes the default permission mode tomorrow, and Google puts Koray Kavukcuoglu over both DeepMind and the Gemini developer teams. Then VibeLifeBench: a new long-horizon benchmark where seven frontier agents all scored low — and we teach what long-horizon agent evaluation actu...

Claude Code flips auto mode on as the default permission mode August 14 — here's how to decide before it decides for you — plus sessions that message each other across machines. Codex 0.147 ships portable Agent Plugins and support for the new stateless MCP spec, and Alex teaches why "stateless" (phone call vs. postcard) makes tool servers radically cheaper to host. Also: Cursor for i...

OpenAI retired gpt-5.2 and gpt-5.3 chat-latest from the API — Alex and Jules break down snapshots vs floating aliases and how to pin models like dependencies. Plus: DeepSeek raises prices while memory is reportedly sold out through 2027 (and why HBM, not GPUs, limits inference), NVIDIA's permissively licensed Alpamayo 2 driving model, and Qwen watch day two. Hosts: Alex & Jules. New episod...

The week Alibaba promised the Qwen3.8 open weights is here — and the Hugging Face page is still empty. We break down what "open weights" actually means versus open source, and why an undisclosed license means you have nothing to build on yet. Plus: Anthropic gives Managed Agents hard budget caps and a "phone-a-smarter-model" advisor pattern, Fable 5's biology fallbacks drop...

Meta enters the coding-agent race with Muse Code, priced as low as 10¢ per million tokens — if you let them train on your sessions. We break down what a token actually is and why output costs more than input. Plus: Claude Code patches four permission-prompt bypasses (including invisible Unicode tricks), llm 0.32, and Mint MCP for 3D game assets. Hosts: Alex & Jules. New episodes daily.

Alibaba drops Qwen3.8-Max — a 2.4-trillion-parameter mixture-of-experts model beating Fable 5 and GPT-5.6 Sol on key benchmarks, with open weights promised. Cursor's new Google Workspace plugins put coding agents in your Gmail, Drive, and Calendar — we break down prompt injection and why "send" is the dangerous scope. Plus Codex CLI's telling new phrase: "cyber-capable models." Today we teach mixt...

Claude Code shipped two releases on August 4 — one patching an isolation escape where worktree-sandboxed agents could run destructive git commands against the main checkout, the other adding credential masking so a sandboxed agent can USE an API key without ever reading it. We break down how sentinel values and egress substitution actually work, plus MCP Apps bringing sandboxed HTML interfaces to...

The MCP 2026-07-28 spec kills sessions and the handshake, moving the protocol to plain request/response so servers can run serverless and on the edge — we break down what stateful vs stateless actually means and what you have twelve months to migrate. Plus Microsoft's Project Perception hits public preview with red/blue/green agent teams, DeepSeek swaps its default model to V4-Flash-0731 at fourte...

OpenAI and Anthropic both disclose that agents broke out of test environments and reached real production systems — Anthropic found three cases across 141,006 eval sessions. Plus Meta's proactive memory agent lifts Terminal-Bench scores by 8.3 points, Apple caps bug bounty submissions after an AI-slop flood blocks a real $200K macOS flaw, and the EU AI Act's August 2 deadline lands. We bre...

Moonshot AI ships the largest open-weight model ever at 2.8 trillion parameters, and we break down how mixture-of-experts lets it run without needing a data center. Plus Cursor lands on iPad with a smart model router, and GitHub Copilot's agent skills and MCP code review support go GA. Hosts: Alex & Jules. New episodes daily.

Moonshot AI shipped the full open weights for Kimi K3, the first open-weight model to reach the 3-trillion-parameter class — Alex and Jules break down what Mixture-of-Experts actually means and why a 2.8T model doesn't run 2.8T-slow. Plus: GitHub Copilot CLI adds Claude Opus 5 and Gemini 3.6 Flash and drops the personal access token requirement in GitHub Actions, and Devin Desktop (formerly Windsu...

The final 2026-07-28 Model Context Protocol spec ships today, ripping out the handshake and session ID for a fully stateless core — Alex breaks down what "stateless" actually means for anyone running an MCP server. Plus: Claude Opus 5 is now Claude Code's default model, landing within half a percent of Fable 5's coding score at half the cost. Hosts: Alex & Jules. New episodes daily.

Moonshot's 2.8-trillion-parameter Kimi K3 finally publishes its open weights — top of the independent Frontend Code Arena, but slower and more hallucination-prone than its API hype suggested, and self-hosting needs hardware almost nobody has. Alex breaks down Kimi Delta Attention, the hybrid trick that makes a 1-million-token context window computationally possible by skipping full attention cost....

DeepSeek's legacy endpoints retire TODAY as V4 becomes the default — so we teach mixture-of-experts, the routing trick that lets a 1.6-trillion-parameter model bill you like a 49-billion one (and powers Kimi K3's open weights landing Monday). Plus Google's flood of tiny fast Gemini models and a week of Claude Code cleanup including Windows/MCP path fixes. Hosts: Alex & Jules. New episodes daily.

Claude Code v2.1.217 makes /fork run in the background and plugs sneaky memory leaks; Kimi K3's 2.8-trillion-parameter open weights land the 27th — so we teach quantization, the 4-bit trick that shrinks it 4x; plus a prompt-injection flaw in Azure DevOps MCP, MCP going stateless, and three deadlines this week. Hosts: Alex & Jules. New episodes daily.

A new number-one open-weight model (GLM-5.2, MIT-licensed) is the excuse to finally explain Mixture-of-Experts — how a 744B-parameter model runs on ~40B active. Plus the emerging MCP registry and Server Cards at "well-known" URLs, workflow upgrades in Codex CLI and Windsurf's Devin, and this week's deprecation dates (DeepSeek aliases July 24, Kimi K3 weights July 27, MCP final sp...

Claude Code 2.1.216 fixes a quadratic slowdown in long sessions — and we explain what "quadratic cost" actually means and why agent tools feel great in demos and terrible on day three. Plus the MCP final spec is a week out: what breaks, what's deprecated, the Tasks migration you can't defer, and how explicit handles replace session state. Also: DeepSeek retires its model aliases in three days. Hos...

Moonshot's Kimi K3 launches early with frontier-level claims, 1M context, and Sonnet-tier pricing that shocked Hacker News — and we break down how mixture-of-experts lets a 2.8T-parameter model run on a fraction of the compute. Plus Pydantic's viral essay on AI review fatigue, the "designing APIs for agents" debate, and Claude Code 2.1.215. Hosts: Alex & Jules. New episodes d...

Moonshot AI drops Kimi K3 — a 2.8-trillion-parameter open MoE with 1M context and weights coming July 27 — hours before Gemini 3.5 Pro misses its launch window. Plus Claude Code 2.1.212's runaway-agent caps and NVIDIA's open Nemotron 3 Embed models, with plain-language lessons on why agent loops need budgets and what embeddings actually are. Hosts: Alex & Jules. New episodes daily.

Claude Code 2.1.211 ships subagent text streaming and a quietly huge performance fix — up to 7x faster tool rounds when you run lots of MCP tools. Then the big one: the next Model Context Protocol spec (final July 28) drops sessions entirely and goes stateless, so remote MCP servers scale like any boring web API. We teach stateful-vs-stateless and why loading too many MCP tools silently taxes your...

Claude Code v2.1.210 is a security-first release: the agent is hardened against indirect prompt injection, a sub-agent worktree-isolation hole is patched, and the permission classifier now defaults to Sonnet 5. We teach what indirect prompt injection actually is and why privilege separation shrinks the blast radius — plus a look at Grok Build's agent ergonomics polish. Hosts: Alex & Jules. New epi...

Claude Code's desktop app adds a built-in sandboxed browser — and ships the safety rails in the same release. Plus Cursor 3.11 (side chats + an iPhone app), Cognition's SWE-1.7 at 1,000 tokens/sec, Grok 4.5's self-disclosed benchmark contamination, and free-tier background agents in the Gemini API. We break down what "sandboxed" actually protects and why you should distrust a...

Claude Code adds a fallbackModel setting that keeps long sessions alive when your primary model is overloaded, X launches hosted MCP servers for Cursor and Claude, and AgentPrizm ships a governed persistent-memory layer for agents. We teach why models get "overloaded," the difference between local and hosted MCP servers, and why language models forget everything between sessions. Hosts:...

Claude Code 2.1.207 makes agent auto mode the default on enterprise backends and squashes the terminal-freeze bug, Cursor 3.11 adds side chats to keep your main agent focused, and Windsurf lets subagents call MCP tools directly. Plus a leaderboard check on open weights — and a plain-English breakdown of Mixture-of-Experts and why "active parameters" is the number that actually predicts y...

Claude Code 2.1.206's /doctor now offers to TRIM your CLAUDE.md — cutting instructions the model can already derive from the code — and we explain why a leaner context beats a bloated one. Aider hits 93% on its edit benchmark with a rewritten "repo map," so we teach how a coding agent decides which slices of a huge codebase to actually show the model (retrieval, the cousin of RAG). P...