New tools, libraries, and updates worth knowing — curated.
What skillmeld composes now installs into Claude Code, Codex, Cursor, Gemini CLI, Copilot and the other agents that read the shared skills folder, in one command; it also ships as an Agent Plugins package, and every skill carries a verdict on how it behaves outside Claude Code. The skill that drives it runs from any of those agents too.
TypeSafe AI's Jev takes a block of state and typed questions (yes/no, a choice among up to 255 options, a score on a rubric) and returns calibrated probabilities in about 100 ms, at $0.042 per million input tokens with output free. Two weeks after launch, independent runs put it a few points behind Opus on classification at a twentieth of the latency.
The second model in the 5.5 family: more than 30% faster output, up to 30% less per task, the same $2 / $10 per million tokens, and 70.6% on Terminal-Bench 4.0 against Sonnet 5's 10.3%. Two points below Opus 5.5 on GDPval-AA.
One-line install from PyPI, the Claude Code plugin added straight from GitHub, a second scout over curated awesome-lists, and a sharper gate: the scripts a skill ships are named, repeated hosts collapse to one finding, SkillSpector findings carry their explanations, and folded descriptions and hyphenated names are discoverable again.
Every Grasshopper definition can now carry a browser page the plugin serves on your own machine, in step with the canvas both ways: sliders and panels push into Grasshopper, every solve pushes back into a 3D view, drawings and figures, and one button bakes a frozen HTML report. Using it needs no Claude session.
The third .ify tool: 999 simulated designs of a robotically assembled CLT house, scored on carbon, robot time, and how comfortably a robotic arm places the panels. Live Pareto fronts, priority re-ranking, parallel axes, and every render, no install.
Four new core rules close gaps a taxonomy coverage benchmark surfaced, NVIDIA SkillSpector joins the optional scanners, the weekly catalog build detects licenses with scancode, and a new scout searches the SkillsMP registry for candidates.
The hosted catalog is live: a signed, hash-pinned index of community skills rebuilt weekly, so discover and fetch work with zero setup. Plus hardened Skills-API emit and a security bump.
The second .ify tool is on the Rhino Package Manager: stage inputs on a Wireify socket, say what the component should do, and Claude converts it in place into a stock Python 3 component. Saved files carry no plugin dependency.
A drop-in replacement for Sonnet 4.6 at the same price, with its biggest gains in coding and agentic work, adaptive thinking on by default, and a 1M-token context window.
An MIT-licensed agent from Nous Research that creates and refines its own skills from experience, runs across Telegram, Discord, Slack, and the CLI, and works with any LLM provider.
A new model family — Sol (most capable), Terra (a cheaper balanced option), and Luna (fastest and lowest-cost) — out first as a limited preview to vetted partners through the API and Codex.
A code-first CAD framework on the OpenCASCADE kernel, with a clean Pythonic API for precise, manufacturable models.
Invite Claude into a channel and @-tag it to hand off work — it runs asynchronously, keeps channel context, and follows up on its own.
The IfcOpenShell-based add-on (formerly BlenderBIM) edits real IFC data directly inside Blender — no lossy round-trips.
A community skill package spanning Blender, Bonsai, IfcOpenShell, and Sverchok — aimed at reliable, repeatable AEC Python code.
A year on, MCP is the common way agents reach tools and data — adopted across the major vendors, with thousands of servers in production.
Why packaging procedural know-how as loadable skills beats stuffing it all into the prompt — and how progressive disclosure keeps context cheap.
The essay behind a lot of current practice — simple, composable patterns over heavy frameworks, with orchestrator-workers as the workhorse.