Meta’s Sept 15 WhatsApp Business Tools MCP lets Claude, Cursor, Codex, or ChatGPT create Business accounts, verify numbers, register Cloud API access, and draft messaging templates. It expands Meta’s MCP lineup beyond ads and Social Technologies.
Since Oct 30, the computer use tool is GA in both the Interactions and GenerateContent APIs: your loop sends a screenshot, gets back a UI action with an intent explaining it, executes it (e.g. with Playwright) and repeats. Responses can carry a safety decision of require confirmation or blocked, and Google still says to run it in a sandboxed VM or container (docs)…
Released this week under Apache 2.0, it sits between your agent and its model server on 127.0.0.1:8787 , blocking calls to undeclared tools or to hosts nobody mentioned in the conversation and running about 110 atif scan detectors for things like reading /proc/1/environ or curl noproxy , with Claude Code, OpenCode, Pi and three other harnesses tested in CI. Start with log only , and note it can't see URLs a script builds at runtime, so keep default deny egress in your sandbox (Tensor Journal)…
Released Oct 29 on Hugging Face in FP8 and NVFP4 (plus GGUFs), the 1.05T total, 49B active multimodal model has 1M context and day 0 support in vLLM, SGLang, Transformers and llama.cpp, and is also in Mistral's Vibe Code agent. Its modified MIT license allows commercial use except Model as a Service hosting, and the API costs $0.70/M input and $2.10/M output (Hugging Face)…
Launched Oct 30 under Apache 2.0, it runs as an editor plugin for Unreal Engine 5, Unity 6 and Godot 4 on Windows and macOS, plays your game with real inputs from prompts like "verify the tree shakes when you hit it with an axe", writes a report, and learns each game as it goes. It takes your own API key, or autopilot connect claude / autopilot connect codex lets Claude Code or Codex drive it on your existing subscription (GitHub)…
The Apache 2.0 plugin from Microsoft Research and UC Berkeley works with Claude Code, Codex, Copilot CLI and Pi: it prunes tools a repo never uses, redirects turns stuck in failure loops, and can optionally inject learned repo facts, cutting mean SWE Bench cost 18.4–30.9% in the paper's runs. It needs Python 3.10+ and no API key of its own, and everything it learns stays in /.tokeneconomizer (MarkTechPost)…
Spotted something we missed? Start a thread.
