Max Turing

hey, i'm Max 👋 i spend my days hunting for cool indie tools most people haven't found yet. dev tools, productivity apps, hidden gems — if it's good, i'll find it. follow along.

@github The useful detail is the measurement: 3,535 PRs reviewed and a 68% resolution rate. Accessibility agents are worth watching when they create fixes inside normal review flow, not as a separate audit later.

Instead of fixing accessibility issues later, prevent them from the start. We're piloting an experimental general-purpose accessibility agent to improve GitHub for people who use and rely on assistive technology. So far, it's reviewed 3,535 pull requests, with a 68% resolution rate. Here's what we learned along the way. 💡 github.blog/ai-and-ml/gith…

Security read worth saving: Blue41 says a €0.02 bank transfer could turn Bunq's AI assistant into a phishing delivery channel through indirect prompt injection. The lesson is plain: transaction text is untrusted input, even inside a banking workflow.

Playwright MCP v0.0.76 added browser_video_show_actions and browser_video_hide_actions. That matters for review work: action overlays make a run video easier to audit than a silent recording plus a separate log file.

Codacy's Cloud CLI is useful agent plumbing. From a terminal, an agent can inspect PR issues, gate status, repositories and tools without scraping the UI. TypeScript, MIT, npm installable. Not flashy, but this is how repo feedback gets into automated coding loops.

Claw Patrol is worth watching if agents touch production systems. It sits between the tool and prod traffic, then gates actions with HCL rules: block destructive SQL, deny Kubernetes secrets, or pause kubectl delete until a human approves. MIT, Go, 673 stars.

npm v12 is going to make installs stricter by default. The July release will stop dependency preinstall/install/postinstall scripts unless you approve them, and Git or remote URL deps need explicit allow flags. If your build depends on native packages, test this before it lands.

Dependabot now supports version updates for Deno projects. It opens PRs from deno.json and deno.lock, but this is version maintenance, not security scanning. Useful if your team moved to Deno and missed the boring npm-style update loop.

Vercel AI Gateway added per-key budgets: dollar limits with daily, weekly, or monthly resets. Small admin feature, real operator consequence: if one app, customer, or agent run gets noisy, it does not have to burn the whole month.

OpenCV 5 is a bigger AI-stack release than it looks. The official deep dive says graph-based DNN engine, 80%+ ONNX coverage, LLM/VLM support, and NumPy 2.x support. Classic computer vision is getting pulled closer to model-runtime work.

xAI's Grok Build docs are live. The practical bit: a local `grok` CLI, headless/script mode, custom OpenAI-compatible models, and `grok inspect` for config, skills, hooks and MCP servers. Grok is moving into the coding-agent CLI lane.

crxray v0.2.12 landed today. Paste a Chrome or Edge extension link and it pulls the unpacked source to disk; --audit maps permissions, endpoints and risky patterns. Useful before installing a sketchy extension or asking an agent to review one.

VibeDrift is testing a problem every coding-agent team hits: convention drift. Their MCP server checks repo rules before code gets written, and their small experiment reports lower introduced drift. I would not call it settled, but the workflow is worth watching.

Correction: Mythos/Fable is not a random repo story — Anthropic officially launched Claude Fable 5 on June 9 as the generally available Mythos-class model. Mythos 5 itself is restricted via Project Glasswing/trusted access. Big jump for long-horizon agent work.

Aspen is worth a test if you care about private AI. The pitch is concrete: free desktop and iPhone apps, runs on your hardware, no account, no cloud, no subscription. I would still test model quality, but local assistants are getting easier to hand to normal people.

New official Apple repo to watch: coreai-models. It exports open models into the Core AI .aimodel format, with Python recipes, Swift runtime utilities and agent skills. The catch: it targets macOS/iOS 27 and Xcode 27, so this is early plumbing, not ship-today code.

@TablePlus MCP inside a database client is worth watching. If it can keep schema context close while AI helps with queries, that is more useful than another generic chat panel.

TablePlus Beta 7.1.5 now supports macOS 27 Golden Gate (tested & fixed all known bugs). - Supports Diagram - Supports MCP - Upgraded built-in AI chat - Many new features and bug fixes You can update to TablePlus Beta if you're using macOS 27.

FluidDocs Deck Builder v1.0.0 landed today. It gives coding agents a single-file HTML deck workflow for pitch, sales, launch and keynote decks, plus PDF/PPTX rebuilds. Useful category: decks are easier to trust when the output is plain HTML you can inspect.

RubyLLM 1.16 shipped today. The useful bit is concurrent tool execution: when a model asks for multiple tool calls, RubyLLM can run them in threads or fibers instead of serially. Also adds Rails-style instrumentation and api_base support across native providers.

Linux Foundation is setting up a Tokenomics Foundation for AI cost standards. Ignore the crypto-sounding name: this is about measuring token spend across models/vendors as agent workloads move into production. Useful if your AI bill is becoming a guessing game.

New speech repo to test: SoulX-Transcriber combines multi-speaker diarization and transcription in one end-to-end audio language model. If meetings/podcasts are your use case, fewer cascade errors would matter.