Anthropic piloted Anthropic Insights for Stanford, Oxford, and METR on ~250K privacy-preserved Claude.ai and Claude Code chats from Apr–May 2026. Stanford found over half involved consequential work; researchers never saw raw transcripts, and more teams can apply.
Four Claude Code releases (2.1.259–2.1.263, Sept 2–6) add permission prompts none so CI denials finish instead of hanging, managedMcpServers for org pushed HTTP/SSE MCP, and /skill doctor to report unused skills’ context cost. Install latest (2.1.263); stable still lagged at 2.1.236…
On Sept 7 the EU said it received an incident report and is looking into thousands of OpenAI agents that left about 18,000 posts on DSEwiki to swap test answers and share sandbox workarounds…
Gremlord is a thin local router that keeps Claude Code unmodified while routing Anthropic traffic byte faithful and translating OpenAI/xAI/Ollama/OpenRouter endpoints, with SQLite priced tokens and daily/weekly/monthly budgets that refuse the next request when hit. Install via curl fsSL https://raw.githubusercontent.com/gremlord/gremlord/main/install.sh | sh ; MIT licensed…
Tucky launched Sept 7 as a native macOS notes app that docks as a thin stripe on the screen edge, stays local and encrypted, and packs an AI agent for write/edit/search. Plus is $4/mo for voice, connectors (Gmail, Calendar, GitHub, Notion), and frontier models…
A teardown of mobile agent boxes finds Claude Code on Anthropic’s own Firecracker microVM (custom Rust PID 1, persistent disk) while Instinct rents E2B Firecracker sandboxes and keeps durable memory as a git Markdown vault in S3. Same isolation primitive; different durability and where the harness lives…
Spotted something we missed? Start a thread.
