🦥 GitHub Got Smarter, Java Got Better

Hello friends!

Welcome to this week’s Sloth Bytes. I hope you had a great week.

Granola is an absolute must if you’re in a lot of meetings.

It silently transcribes your meetings in the background. There’s no bot. No "recording in progress." Nobody knows it's there.

It works on Zoom, Google Meet, Teams, or in-person. When the meeting ends, you get a clean summary, action items, and a chat interface to ask questions from any past meeting.

Connect it via MCP and that context travels straight into your other tools.

Free for a month with code CODINGSLOTH.

The Supabase team has been on a roll lately and this is one of the weirder (and coolest) things they shipped.

supabase.sh lets you browse the entire Supabase documentation over SSH: yes, from your terminal, in bash. The immediate use case? Agents. You can pipe it straight into Claude Code:

And your AI coding agent now has access to up-to-date Supabase docs without hallucinating outdated API syntax. It's part of a broader push Supabase has been making toward AI-friendly developer tooling. They also recently shipped 30 Postgres best practices specifically written for AI agents, and became an official Claude connector.

This pairs nicely with llms.txt, a proposed standard (kind of like robots.txt but for LLMs) where websites expose a clean markdown file at /llms.txt that AI tools can read at inference time. The idea is simple: your website was built for humans, LLMs need something different. Give them a structured, concise entry point. A bunch of major projects are already adopting it.

Why it matters: Docs-over-SSH is a fun party trick, but what it signals is more interesting. The dev tooling world is actively rethinking how documentation gets consumed. When AI agents are doing the reading, the formats need to change. Expect more of this.

Java 26 officially dropped on March 17, 2026. It's a non-LTS release with 10 JEPs (JDK Enhancement Proposals), and before you close this… hear me out, because a few of these are genuinely neat.

Here are the highlights:

It's a non-LTS release, so you have 6 months of support before they nudge you toward Java 27. If you're on Java 21 LTS, no rush… but it's worth watching what's graduating to stable.

Why it matters: Java has been quietly getting more expressive. Pattern matching, structured concurrency, lazy constants: these are features that make Java feel less like filing your taxes and more like a language people actually enjoy writing. The JVM ecosystem isn't going anywhere.

This one flew a little under the radar last week but it's a big deal.

GitHub Copilot CLI now lets you plug in your own model provider instead of being forced to use GitHub's hosted routing. We're talking:

There are some catches. Your model needs to support tool calling and streaming, and they recommend a 128k+ context window for best results. But the fact that you get the full agentic terminal experience (explore, task, code-review sub-agents) with whatever model you want? That's a legitimately good feature.

Why it matters: The "which AI tool should I use" conversation is increasingly becoming "which model should I route through which tool." BYOK support is table stakes now, and GitHub just delivered it for the CLI. If you're already paying for Anthropic or Azure OpenAI credits, you can now use those in Copilot CLI instead of paying twice.

Quick explainer for newer devs: a "stacked PR" is when you break a big change into a series of smaller pull requests that each build on the one before. Instead of one giant PR with 47 files changed (that reviewers will absolutely not read carefully), you get a chain of focused, reviewable slices.

This is a workflow a lot of senior engineers swear by, and until now you had to use third-party tools like Graphite or manage it manually. GitHub is now building it natively with GitHub Stacked PRs and a gh stack CLI.

Here's what it actually does:

There's also AI agent integration: run npx skills add github/gh-stack to teach your coding agent how to work with stacks and break up large diffs automatically.

One catch: it's currently in private preview with a waitlist. But the docs are already public and worth reading if you want to understand the model.

Why it matters: Large PRs are one of the most common sources of review bottlenecks and merge conflicts on teams. Stacked PRs are how a lot of high-output engineers already work, and having it native in GitHub (instead of requiring a separate tool everyone has to agree to install) removes a lot of the friction.

ELF & Dynamic Linking deep dive -the article that reminded me I actually love low-level systems stuff; a proper breakdown of what happens between ./app and your program actually running.

Survey on LLM Inference Engines - a deep academic paper comparing 25 open-source and commercial inference engines for running LLMs. Heavy reading but useful if you're thinking about deploying your own model. (Long read, but very interesting!)

What is robots.txt? - a file that's been quietly running the internet since 1994 that every developer should know about. It's also the inspiration behind the newer llms.txt standard for AI agents.

Mechanical Sympathy - a sharp, honest piece on why AI coding agents still lack something every great engineer has: the instinct to work with the system instead of against it.

How Cloudflare Deploys MCP Across Their Whole Company - Cloudflare published exactly how they rolled MCP out company-wide: the security architecture, the cost cuts, the pitfalls. Worth reading before you wire up AI agents to anything important.

Anthropic's Guide to Securing Your Systems as AI Makes Hacking Faster -Anthropic's security team published a practical, no-fluff guide on what actually changes when AI makes finding and exploiting vulnerabilities dramatically faster.

OpenScreen - a free, open-source alternative to Screen Studio (which costs $29/month) for making polished screen recordings with zoom effects, annotations, and custom backgrounds. MIT licensed, works on Mac and Windows, no watermarks.

delphi.tools - a collection of small, useful design and dev tools (QR generators, color converters, SVG optimizers, regex testers, favicon generators...) with zero logins, zero tracking, and zero nonsense. A breath of fresh air.

Ballerine - open source KYC (Know Your Customer) and KYB (Know Your Business) infrastructure. If you ever need to add identity verification to a product, this is where to start instead of building it yourself.

pytesseract - a Python wrapper for Google's Tesseract OCR engine that lets you extract text from images in a few lines of code.

That’s all from me!

Have a great week, be safe, make good choices, and have fun coding.

If I made a mistake or you have any questions, feel free to comment below or reply to the email!

See you all next week.

If your company is interested in reaching an audience of developers and programming enthusiasts, you may want to advertise with us here.