Local AI for coding, and what to check before connecting it
Magnitude offers a way to run coding AI on your own computer, though its speed claims need testing. We also look at Cursor's business connectors, a Google model that isn't widely available yet, and a retailer's reported gains from AI shelf checks.
The Big Story
Google Gemini Blog
Gemini 4 Argon: our next era of frontier intelligence
Google announced Gemini 4 Argon for coding, knowledge work, and cybersecurity, but access is currently limited to trusted testers and selected cyber defenders. It plans to open access later, starting with paid API customers and Google AI Ultra subscribers, without giving a date. The introductory API price will be $2 per million input tokens and $10 per million output tokens (cached input gets a 95% discount); Google says the model can produce up to 1 million output tokens in one run.
Don’t plan a workflow around Argon yet. Wait for broader access, then compare its cost and results on your own tasks rather than relying on benchmark claims.
For Everyone
TechCrunch AI
Hearing tech startup Legato launches its AI hearing glasses
Legato has launched its AI-powered hearing glasses in the U.S., starting at $999 for adults with up to moderate hearing loss. The company says Legato Frames separate speech from background noise and process sound on the glasses without cameras or recording. They support prescription lenses, offer 10–12 hours of battery life, and are sold online or through select eye care providers (where they may qualify for vision insurance coverage).
If hearing aids are uncomfortable or impractical, compare the frames with conventional options and check insurance coverage before paying the $999 starting price.
For Builders
AI TLDR
Magnitude — a local inference engine for agents, up to 2x faster than llama.cpp
Magnitude is a free, open-source desktop app for running AI models locally and connecting them to coding agents such as Codex, Claude Code, and Pi. It supports macOS, Windows, and Linux, including machines without a GPU, and provides an OpenAI-compatible API. Its developers claim up to 2× faster inference than llama.cpp on some hardware; those figures come from the project’s own benchmarks. They say prompts and files stay on the machine after a model is downloaded.
If your team handles sensitive code or pays significant model API fees, test a suitable local model against your existing workflow, checking quality and hardware costs as well as speed.
GitHub Trending
plugins: Cursor plugin specification and official plugins
Cursor's public plugins repository provides a marketplace manifest and individual plugin definitions for its AI coding environment. The catalog includes developer tools and connectors for Google Workspace, Microsoft Teams, Salesforce, HubSpot, Shopify, Xero and other services. A plugin can package agent instructions, rules and MCP server configuration (a way for the assistant to connect to external tools); the repository is MIT licensed, and listings include both Cursor and outside authors. The README describes what's available but doesn't establish pricing or the permissions required by each connector.
If your team already uses Cursor, check an individual plugin's author and requested access before connecting business accounts, especially email, customer records or financial systems.
AI TLDR
Pi 0.99 — the minimal coding agent adds MCP and a codemode sandbox
Pi now supports MCP (a standard way to connect AI agents with outside tools) in its core. Its Codemode sandbox lets an agent use JavaScript to coordinate multiple tool calls, and loads automatically when MCP is configured; it can also be enabled on its own. Earendil says this helps with tool discovery and composition, while acknowledging that many MCP servers still return unstructured text that limits the approach.
If your team already uses Pi and MCP tools, try a small workflow that combines two tools and check the results and permissions before making it part of daily operations. Others don’t need to change their setup.
For Business
AI News
Vinasoy reports 20% out-of-stock reduction with AWS AI
Vinasoy says its photo-based shelf monitoring system, built with Renova Cloud on AWS, cut its out-of-stock rate by 20% within two months. Sales staff now receive display scores in one to two days instead of waiting more than 20 days for a manual review, and monthly coverage rose from ~17% to more than 70% of outlets. Those are company-reported results; the article gives no accuracy figure for the AI scoring.
If store shelves are a recurring blind spot, first measure how long manual checks take and how often products run out, then test photo-based monitoring on a small set of stores before investing in a custom system.