AI library

Weekly summary

Week 2026-W43

25 October 2026

The week in AI

  • Agent Skills gets neutral governance: Anthropic donated the spec, name, and agentskills.io to the Agentic AI Foundation (Linux Foundation), which also governs MCP. Maintainers from Anthropic, Google, OpenAI, and others now steer it, and Anthropic counts over 70 platforms that load skills. (source)
  • Code mode for any MCP server: MCPtoCodeMode (MIT) hides all your MCP servers behind search, describe, and execute, and runs the model's JavaScript in a QuickJS sandbox. Add it with claude mcp add codemode -- npx -y mcptocodemode serve. (source)
  • Claude API version 2026-10-21: the first new API version since 2023-06-01. It removes temperature, top_p, and top_k (400 on every model), and the new SDK majors send keys as Authorization: Bearer and move to Apache 2.0. Claude Code's /claude-api api-version migrates code. (migration guide, 2.1.307)
  • Claude Code sandbox hardening: 2.1.309 closes gaps where shell startup files ran unsandboxed and sandboxed commands could edit the --mcp-config file. If you use Bedrock, Vertex, or a gateway, skip 2.1.304: it broke every request there, and 2.1.305 fixes it. (2.1.308–309, 2.1.305–306)
  • How Anthropic reviews agent-written code: every PR gets a multi-agent review with voting, human approvals scale with risk, and review is capped at three rounds. Merged PRs doubled from June to September 2026 while the reverted share fell by over 40% (Anthropic's own figures). (source)
  • Cheaper speed and longer caches: Opus 5.5 fast mode is GA at $8 / $40 per 1M tokens, and OpenAI's Responses API keeps a prompt cache for 12 hours at 2.5x input to write. OpenAI's Decisions API also went GA, with up to 400 questions per request. (fast mode, 12-hour cache, Decisions)

Rising on GitHub

The report's "Rising" lists are still empty (no snapshot is 7 days old yet), so these come from "New this week". I read each README; all star counts and benchmark claims are the projects' own.

  • morganlinton/BeautifulRust (308 stars): MIT skill and plugin, distilled from five Rust style guides, that makes agents write idiomatic Rust. Install with npx skills add morganlinton/BeautifulRust.
  • EricJujianZou/rederive (192 stars): MIT skill that teaches you a codebase by asking guiding questions, so you work out its design decisions yourself. Install with /plugin marketplace add EricJujianZou/rederive.
  • rextail-ai/lessloc (101 stars): Apache-2.0 plugin that steers Claude Code, Codex, Cursor, and others to the least code that does the job. It claims up to 36% less code and 26% lower cost.
  • kanaworksai-hue/audio-studio-mcp (481 stars): MIT MCP server that generates background music and sound effects locally with Stable Audio 3 (Apple Silicon or NVIDIA). The weights are under the Stability AI Community License.
  • tomtomsatnav/agent-orchestration-standard (76 stars): MIT, read-only, local dashboard for Claude Code sessions and subagents (tokens, cost, file conflicts). It is a macOS-first personal project.

Library maintenance

  • Links: 107 checked, 0 dead, 3 moved, 2 blocked. Three Anthropic articles moved from /engineering/ to /research/, and their url fields are updated: Building effective agents, How we contain Claude, and Redesigning source control. The two blocked links are OpenAI pages that return 403, so no action was needed.
  • Consolidated 10 of the 12 pages updated this week, moving durable facts into the body and keeping Current state to what is recent and still changing:
    • tooling/coding-agents: Current state cut from 13 to 3 entries. New Observability, Context handling, and Add-ons sections hold Ribbon, stalemap, and Flint.
    • tooling/mcp: Current state cut from 13 to 3 entries. New sections cover spec adoption, Claude Code client behaviour, and notable new servers.
    • models/claude: Current state cut from 10 to 4 entries. API version breaking changes, platform governance (inference hooks, user profiles), evaluators, and political neutrality moved into Key concepts.
    • practices/security: Current state cut from 10 to 4 entries. Inference hook details, key handling, safety telemetry, and the nested namespace caveat moved to How to use it; election misuse, Wikimedia, and sandbox fixes moved into Key concepts.
    • frameworks/landscape: nine release entries folded into the framework descriptions and a new Local runtimes (Ollama) entry; three upgrade warnings remain.
    • models/open-weight: Current state cut from 9 to 4 entries. Flint, Voxtral 3, d3, and the Gemma scam detector moved into the body.
    • models/gpt: Current state cut from 9 to 3 entries. New Voice entry; the Ultrafast access disagreement moved into the body.
    • practices/cost-and-caching: Current state cut from 9 to 3 entries. The Ultrafast access disagreement moved into Paying for speed.
    • tooling/skills and agents/tool-use: entries duplicated in the body were removed or folded in.
    • agents/memory and agents/agent-patterns: checked; 3 entries each, no change needed.
  • Added missing Sources links: Ollama 0.40.9 on coding-agents; Claude Code 2.1.305 and Agent Framework 1.23 on security.
  • No stale pages: the oldest last_reviewed date is 2026-10-04 (21 days).
  • Inbox empty. No duplicate tags. No open questions were newly answered beyond those the daily runs had already closed.