Weekly · AI agent toolkit · week 40, 2026
andrej-karpathy-skills + 7 more must-have coding-agent tools · week 40, 2026
This week's must-have repos for AI coding-agent users — andrej-karpathy-skills, markitdown, mcp, and more: Claude Code-first, plus tools that work with any agent, drawn out of the noise.
Claude Code Tool
by multica-ai - A drop-in CLAUDE.md distilling four behavioral guidelines for LLM-assisted coding into Claude Code — a low-friction quick win. Karpathy-inspired, derived from Andrej Karpathy's public notes on LLM coding pitfalls and authored by multica-ai
215,957 stars
View on GitHub →What we said about andrej-karpathy-skills
Alright, this one's interesting precisely because there's no code to speak of. What multica-ai has done is take Andrej Karpathy's public notes on where LLM-assisted coding tends to go sideways and distill them into a single CLAUDE.md file you drop straight into Claude Code. That's it. Four behavioral guidelines that quietly shape how the model works alongside you.
Why does that matter? Because most of us fight the same failure modes daily — the model over-engineering, hallucinating APIs, or plowing ahead without asking. A good CLAUDE.md is basically a system prompt that steers the agent before those problems start. Compared to hand-rolling your own rules or grabbing a bloated agent framework, this is genuinely low-friction: one file, no dependencies.
Is it worth two hundred thousand stars? That's a conversation about how repos trend now. But the underlying idea — treating prompt conventions as shared, versioned artifacts — is where a lot of teams are heading.
We break down more picks like this each week in the newsletter. Link's in the description.
by mensfeld - Give each AI agent its own isolated machine with root, Docker, and systemd. Active defense detects and stops threats automatically
Go · 732 stars
View on GitHub →What we said about coi
So here's a problem you probably haven't thought hard about yet: what happens when your AI agent goes rogue on your actual machine? mensfeld's coi tackles that head-on by giving each agent its own isolated box — full root, Docker, systemd, the works — so if something misbehaves, the blast radius stops at that sandbox. What I find genuinely interesting is the active defense piece. It's not just walls; it's watching behavior and shutting threats down as they happen, which feels a lot closer to how we think about production security than the usual "trust the prompt" approach.
Compared to spinning up throwaway containers by hand or leaning on cloud VMs, this bakes the isolation and monitoring together, and being written in Go means it's lightweight to deploy. If you're running autonomous agents that execute code — especially anything touching your filesystem or network — this is worth a serious look before you get burned.
If you want more finds like this landing in your inbox every week, the newsletter link is down in the description. Go grab it.
Works With Any Agent
🎖️ 🐍 🏠 - Convert many file formats, local or remote, to Markdown for LLMs using MarkItDown
Python · 187,684 stars
View on GitHub →What we said about markitdown
So here's a problem that sounds boring until it wrecks your afternoon: you've got a PDF, a PowerPoint, an Excel sheet, maybe an audio file, and you need all of it in clean Markdown so your LLM can actually read it. MarkItDown from Microsoft is the quiet utility that handles that conversion across a genuinely wide range of formats, local files or remote URLs, with one consistent interface.
What I appreciate here is the focus. This isn't trying to be a document editor or a RAG framework — it's a preprocessing layer, and it does that one job well. Compared to rolling your own with a pile of format-specific parsers, or leaning on something heavier like Unstructured, MarkItDown gives you a lighter, more predictable path to LLM-ready text.
If you're building pipelines, chatbots, or anything that ingests messy real-world documents, this belongs in your toolkit. The Markdown output keeps structure your model can reason about, without the token bloat of raw HTML.
If tools like this are your thing, the newsletter rounds up the best ones every week — link's in the description.
🎖️ #️⃣ ☁️ - Access Azure services including Storage, Cosmos DB and Azure Monitor
C# · 3,722 stars
View on GitHub →What we said about mcp
Let's talk about microsoft/mcp, which is Microsoft's implementation of the Model Context Protocol built specifically for Azure. If you've been watching MCP servers pop up everywhere, this one's interesting because it comes straight from the source — giving your AI agents structured, permissioned access to Azure Storage, Cosmos DB, and Azure Monitor without you hand-rolling a bunch of API glue.
Here's why that matters: most MCP servers you find are community weekend projects that break when an SDK version bumps. This being first-party means it tends to track Azure's own auth model and service changes more reliably. It's C#, so it slots naturally into .NET shops already living in the Microsoft ecosystem.
Compared to writing your own tool-calling wrappers, this saves you the tedious plumbing of exposing cloud resources safely to a model. It's really aimed at teams building agents that need to query logs, read blobs, or hit a database as part of a workflow.
If you want more picks like this every week, the newsletter link is in the description — go grab it.
🎖️ 📇 🏠 - Local code search for coding agents: hybrid keyword and semantic search with SQL relevance ranking over a plain-file index
TypeScript · 31 stars
View on GitHub →What we said about supergrep
Alright, let's talk about supergrep from infino-ai. So the premise here is deceptively simple: it's local code search built specifically for coding agents. And that last part is the key. This isn't grep with a fancy hat — it's hybrid search, blending keyword matching with semantic understanding, then ranking relevance using SQL over a plain-file index.
Why does that matter? When you're wiring up an AI agent to reason about a codebase, feeding it the right context is everything. Vector-database-heavy setups can be overkill — extra infra, extra cost, extra latency. Supergrep keeping the index as plain files is a genuinely pragmatic choice. It's portable, inspectable, and you can commit it or throw it away without spinning up a service.
It's TypeScript, so it slots naturally into Node-based agent tooling. This is early — thirty-one stars — so temper expectations, but the design philosophy is sound for anyone building retrieval into their own agent instead of paying for a black box.
If tools like this are your thing, the newsletter rounds up the best ones each week — link's in the description.
🏠 🍎 🐧 - Headless browser for AI and web automation with a built-in MCP server
Zig · 35,657 stars
View on GitHub →What we said about browser
Okay, let's talk about Lightpanda, because this one caught my attention for a reason that isn't just the star count. It's a headless browser written in Zig, and that choice tells you everything about the intent. Most automation tooling today leans on Chromium under the hood, which means you're shipping a browser engine that was designed for humans looking at pixels. Lightpanda strips that assumption out. It's built for machines — for agents and scrapers that need to parse the DOM and run JavaScript without dragging a full rendering stack behind them. That translates to lower memory and faster cold starts, which actually matters when you're spinning up dozens of parallel sessions.
The built-in MCP server is the tell here. They're aiming squarely at the AI agent workflow, so your model can drive a page directly. Compared to Puppeteer or Playwright, you trade some ecosystem maturity for a much lighter footprint. If you're prototyping agent infrastructure, it's worth a serious look.
If breakdowns like this are your thing, the newsletter link's in the description — go grab it.
📇 ☁️ - Email for AI agents: create inboxes on the fly to send, receive and act on email
TypeScript · 66 stars
View on GitHub →What we said about agentmail-mcp
Here's something clever if you've ever tried giving an AI agent a real inbox. AgentMail's MCP server lets your agent spin up email addresses on the fly, then send, receive, and actually act on messages — all through the Model Context Protocol, so it plugs straight into Claude, Cursor, or whatever MCP-aware client you're running.
Why does that matter? Most agent email setups mean wrestling with SMTP credentials, OAuth scopes, or handing a bot access to your personal Gmail — which nobody sane wants. This flips it: disposable, programmatic inboxes that belong to the agent, not you. Think signup flows, verification codes, or an agent that triages support mail without touching your production account.
Compared to raw Nodemailer wiring or Zapier glue, the MCP approach means the tool definitions come baked in — less prompt babysitting. It's early, sitting around sixty-six stars, and TypeScript-first, so if you're already in that ecosystem it'll feel familiar.
If tools like this are your thing, the newsletter rounds up the best of them every week — link's in the description.
Popular MCP server that enables AI agents to scaffold, build, run and test iOS, macOS, visionOS and watchOS apps or simulators and wired and wireless devices. It has powerful UI-automation capabilities like controlling the simulator, capturing run-time logs, as well as taking screenshots and viewing the accessibility hierarchy
TypeScript · 6,449 stars
View on GitHub →What we said about MobileBuildMCP
Here's one that fills a genuine gap. MobileBuildMCP from Sentry gives your AI agent a real handle on the Apple development loop — not just writing Swift, but actually scaffolding a project, kicking off a build, launching the simulator, and driving it. It can tap through your UI, capture runtime logs, grab screenshots, and read the accessibility hierarchy, which is the part that matters most. That accessibility tree is how an agent understands what's actually on screen, instead of guessing from pixels.
Compared to poking at xcodebuild and simctl through raw shell calls, this wraps the whole toolchain in a clean MCP interface, so Claude or Cursor can reason about state rather than parse terminal noise. It covers iOS, macOS, visionOS, watchOS, plus wired and wireless devices, which is broader than most mobile automation setups.
If you're an Apple platform developer experimenting with agentic workflows, this is worth a weekend. It's TypeScript, so it slots into existing MCP configs easily.
For more picks like this every week, the newsletter link is in the description.