Direct GitHub API reads on 2026-08-23 confirm: mat, The Model Context Protocol's own maintainers publi, A Level1Techs forum deep-dive reproduces concrete

Direct GitHub API reads on 2026-08-23 confirm: mattpocock/skills at 232,450 stars, obra/superpowers at 276,280, and affaan-m/ECC at 242,240 -- all three individually ahead of anthropics/claude-code (142,606) and openai/codex (113,873). This is the first same-session confirmation that three independent third-party skills/plugin-marketplace repositories each hold more stars than the two coding-agent platforms they extend, not a scrape artifact. mattpocock/skills' 24-hour delta (+2,504) against the day's claimed +2,683 gives a fourth consecutive day with a fourth different accuracy ratio for GitHub Trending's daily-delta figure.

The Model Context Protocol's own maintainers published a roadmap naming five priorities, including 'progressive discovery so servers reveal tool catalogs incrementally rather than exposing entire catalogs upfront' -- the first time the protocol itself has named full-catalog-upfront exposure a design problem to fix, rather than something third-party middleware (like the 'token tax' tools this vault tracked from July) works around. The other four priorities: agentic messaging primitives (the Tasks extension, SEP-2663), HTTP-native transport unification, agent identity and enterprise security (DPoP, Workload Identity Federation), and SDK developer experience.

A Level1Techs forum deep-dive reproduces concrete inference-stack hazards behind local LLMs feeling weaker than their benchmarks suggest: different attention-kernel backends shift top-1 token selection by 0.717%-5.831% depending on quantization scheme; int4 KV-cache quantization never recovered from a tool-calling error where int8/BF16 did; NVIDIA's own NVFP4 weight quantization performed worst of five schemes tested (~50% token flips by 88k context); and the same model flipped between correct and incorrect tool calls purely by changing tensor-parallelism degree (TP1 correct, TP2 failed, TP4 correct again), attributed to GPU-to-GPU communication-order effects.

OzBrain launched as a shared knowledge layer that Claude, ChatGPT, Cursor, and other MCP clients connect to via a single URL, reading and writing a conflict-checked, shared 'article' store. This is the third distinct repositioning in five months of the same underlying pattern this vault has tracked since April: a developer-targeted CLI framework (Ar9av/obsidian-wiki, April), then a no-terminal consumer GUI app ('Second Brain', July), and now an explicitly agent-to-agent product where the customer is the agent fleet itself, not a human reader.

Google Play's top-20 shows three AI-labeled titles today (Canva 'AI Photo & Video Editor', Yazio 'AI Calorie Tracker', Filmora 'AI Video Editor&Maker') versus one confirmed the day before, though today's scrape again shows a chart-conflation artifact -- a browser-category list appended after the general top-20, the second such data-quality issue in two days from this source.