AI Daily Dev — October 16, 2026

Anthropic Hands METR Employee-Like Access as First Embedded Nonprofit

Oct 15: Six METR staff get Anthropic hardware, office space, and the same internal risk-assessment tools Anthropic's own safety teams use — Anthropic cannot veto their published findings outside a short list of predefined redaction categories (customer data, legal privilege, model architecture).
Caps a three-tier evaluator ladder: tier 1 is narrowly scoped, tier 2 lets evaluators publish freely with limited info, tier 3 is METR's embedded-employee setup.
Extends the embedded-evaluator program Anthropic launched Sep 18 with Accenture's Faculty unit; SecureBio is now reviewing the CBRN chapter of the August 2026 Risk Report, with Hugging Face in early talks.
Dario Amodei pitched the model in a Sep 12 essay comparing it to onsite bank examiners; Anthropic says it hopes every frontier AI lab adopts a similar arrangement.

industry anthropic.com

Oct 14: Anthropic's fifth APAC office opens alongside an MOU with Singapore's MDDI covering model evaluation, incident response, and SME adoption; GovTech Singapore has 1,500+ engineers using Claude Code to modernize government systems.
DBS runs Claude in software development, GIC is rolling it out to investment teams, Grab's BriX sales assistant saves 40,000 hours/quarter, and Sea's Shopee uses Claude for both customer support and coding.
Singapore ranks second of 121 countries in Claude.ai usage per capita, at ~6x what population alone would predict.
Launch week paired a Builder Summit (300+ developers) with a 700+ ambassador meetup; in-region inference ships on Amazon Bedrock for regulated workloads.

industry anthropic.com

cathrynlavery/diagram-design: 56.1k stars (+10.2k this week) — MIT skill that gives coding agents 44 editorial-grade diagram types output as a single self-contained HTML+SVG file, no runtime deps.
Explicit Mermaid replacement for agent output: flat styling, no shadows, renders standalone in any browser or markdown viewer.
Installs via the open Agent Skills spec into Claude Code, Codex, GitHub Copilot, Cursor, Factory Droid, and Pi.
Diagram catalog grew from 29 types in August to 44 now; weekly #3 trending AI/dev repo behind morluto/rea and mattpocock/skills.

tools github.com

earthtojake/text-to-cad: 20.8k stars (+2.2k this week) — MIT plugin that lets Claude Code, Codex, Cursor, Gemini, and Grok generate parametric 3D models and spit them out as STEP, GLB, STL, or 3MF.
Ships DFM checks, drawing generation, and routed connectors into 3D-print, sheet-metal, and CNC fabrication services from the agent prompt.
Python 3.11+ and Node 20+ under the hood; weekly trending alongside Cursor's own plugin spec and twostraws/SwiftUI-Agent-Skill.
Marks the agent-skills ecosystem spreading past text/code work into real engineering domains with physical output.

tools github.com

mksglu/context-mode: 29.9k stars (+4.0k this week) — Elastic License 2.0 MCP server that sandboxes agent tool output and persists per-session memory in SQLite.
Author-reported 98% reduction in context used by tool calls, with hook-based routing across 17 agent platforms and sandbox execution in 12 languages.
Direct response to the context-pollution problem that drives up token bills on long Claude Code and Codex sessions; third separate 'cut your tokens' repo on this week's trending list after caveman and billion-context.
License is ELv2, not OSI-approved — commercial use allowed, but hosting it as a service for third parties is forbidden.

tools github.com