THX-01 takes a state (ticket, email, agent trace) plus typed questions and returns calibrated answers, numbers, verbatim excerpts, and citations in about 10 ms on one GPU, under Apache 2.0 via pip install thx01. HAL-X says its hosted API now needs no key or sign-up, subject to rate limits.
The Apache 2.0 engine behind LiteRT runs models on phone, web, and desktop GPUs through OpenCL, OpenGL ES, Metal, and WebGPU, adds 5D tensors, and replaces the old TFLite GPU delegate, which gets no new features. The repo at google ai edge/ml drift also ships a SKILL.md so coding agents can write, register, and check custom GPU shaders…
npx create talorys@latest provisions a private Worker, a Pages frontend, and a SQLite backed Durable Object in your account, giving you chat, memory, tasks, notes, and scheduled reminders on Workers AI ( glm 4.7 flash ). It's MIT licensed, single user, has no telemetry, and keeps non AI features working when the daily free AI quota runs out…
Agents can now run vercel domains search , check , price , and buy , guided by a bundled agent skill. Run non interactively, buy returns a structured error with next steps, so the final price, auto renew, and registrant confirmation stays with you…
MCP servers can now set meta: {"antigravity/alwaysLoad": true} on a tool so the agent gets it directly instead of discovering it on demand, and plugins can declare subagents in a root level agents.json . The macOS sandbox also now blocks commands from reading /Library/Application Support , where desktop apps keep credentials, unless you allow it…
A new Fix item in JetBrains diagnostic menus hands the problem to Copilot in inline chat (agent mode when available), and a new setting stops MCP servers for Copilot and Claude from starting automatically. Enterprise admins can also set the default agent model, and the plugin now requires JetBrains IDE 2025.2 or later…
Spotted something we missed? Start a thread.