Weekly · AI agent toolkit · week 44, 2026
andrej-karpathy-skills + 7 more must-have coding-agent tools · week 44, 2026
This week's must-have repos for AI coding-agent users — andrej-karpathy-skills, markitdown, clawdcursor, 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
225,655 stars
View on GitHub →What we said about andrej-karpathy-skills
This is a single configuration file you drop into Claude Code that nudges the model toward more disciplined coding behavior. It's called andrej-karpathy-skills, and the whole thing is essentially a distilled CLAUDE.md built around four guidelines pulled from Karpathy's public commentary on where LLMs tend to trip up when they write software. What I like here is the restraint. There's no framework to install, no new dependency tree — you're just giving the assistant better guardrails about things like over-engineering, hallucinated APIs, and sprawling diffs it can't reason about. For anyone already living in Claude Code day to day, that's a near-zero-cost experiment. Compared to the heavier agent-orchestration tools making the rounds, this sits at the opposite end: it's a prompt-discipline artifact, not a system. The caveat is that behavioral guidelines in a config file are suggestions, not enforcement, so results will vary with task complexity. Treat it as a sensible baseline you then tune to your own codebase, and it earns its place in the repo.
Works With Any Agent
🎖️ 🐍 🏠 - Convert many file formats, local or remote, to Markdown for LLMs using MarkItDown
Python · 192,625 stars
View on GitHub →What we said about markitdown
This one takes the messy pile of formats you deal with every day—PDFs, Word docs, PowerPoint slides, Excel sheets, even images and audio—and flattens them into clean Markdown your model can actually read. It's called MarkItDown, from Microsoft, and the reason it's resonating is pretty practical: feeding raw files into an LLM is a nightmare, but Markdown preserves structure like headings, lists, and tables without the token bloat of HTML or the chaos of raw PDF extraction.
What I like is the focus. This isn't trying to be a full document database or a RAG framework—it's the unglamorous conversion layer that sits in front of all that. Compared to something like Unstructured, it leans lighter and simpler, which makes it a solid default when you just need text in, Markdown out.
If you're building ingestion pipelines, document chatbots, or anything that preprocesses human files for a model, it's worth a look. One thing to test early: how it handles your specific PDFs, since quality there varies wildly across tools.
📇 🏠 🍎 🪟 🐧 - Desktop control for any AI agent on Windows, macOS and Linux. Reads the screen as text (accessibility tree, then OCR) and acts on controls by name, with screenshots only as a last resort, so a task costs a fraction of the tokens of a screenshot loop. 7 compound tools, one safety gate, verifies its own actions, and pauses for the user at sign-ins. Local-only, any tool-calling model
TypeScript · 406 stars
View on GitHub →What we said about clawdcursor
This gives any AI agent the ability to actually operate your desktop—clicking buttons, filling fields, navigating apps—across Windows, Mac, and Linux. It's called clawdcursor, and the clever part is how it reads the screen. Instead of screenshotting everything and burning tokens on image parsing, it reads the accessibility tree as text first, falls back to OCR, and only reaches for a screenshot when it genuinely has to. That ordering is the whole thesis—it acts on controls by name, so a task costs a fraction of what a screenshot loop would.
Most agent-control tools lean hard on vision models, which is expensive and slow. This flips the priority and gives you seven compound tools, a single safety gate, self-verification of its own actions, and—crucially—it pauses and hands control back to you at sign-in screens. That last bit shows someone thought about real-world use, not just demos.
It's local-only and works with any tool-calling model, so you're not locked into one provider. If you've wanted to wire an agent into your actual workflow without the token bleed, this is worth a look.
🎖️ 📇 🏠 🍎 🪟 🐧 - Free, open-source screen recorder and video editor that agents drive over stdio (the desktop app's own executable started with mcp): record the screen, trim, zoom, caption, annotate and export demo videos, and add custom WGSL shader effects that export exactly as previewed
TypeScript · 5,015 stars
View on GitHub →What we said about openscreen
Imagine recording your screen, trimming the dead air, adding zooms and captions, then exporting a polished demo — all without touching the editor yourself. That's the pitch behind openscreen, a free, open-source screen recorder and video editor where the desktop app exposes its own executable as an MCP server. So an AI agent can drive the whole pipeline over stdio: start a recording, annotate it, apply WGSL shader effects, and export exactly what the preview showed.
What makes this interesting isn't the recorder itself — tools like OBS or Screen Studio already cover that ground. It's that the editing surface is fully scriptable by an agent, which turns repetitive demo-making into something you can automate. If you're building developer tooling, writing release notes with GIFs, or churning out product walkthroughs, that's a real time sink gone.
The WGSL support is a nice touch too — custom GPU effects that render deterministically on export, so no surprises between preview and final file. It's cross-platform, TypeScript under the hood, and worth a look if your demos pile up faster than you can edit them.
🏎️ 🏠 🎖️ - Official Nhost MCP server, bundled with the Nhost CLI. Lets AI coding assistants read your GraphQL schema, run the queries and mutations you allow, manage permissions and migrations, update Nhost Cloud project config, and search the docs
TypeScript · 9,370 stars
View on GitHub →What we said about nhost
Imagine your AI coding assistant actually understanding your backend — not guessing at your database shape, but reading the real GraphQL schema and running the queries you explicitly permit. That's what nhost/nhost brings to the table with its official MCP server, bundled right into the Nhost CLI. For the unfamiliar, Nhost is an open-source Firebase alternative built on Postgres, Hasura, and GraphQL, and this release wires it directly into tools like Claude or Cursor through the Model Context Protocol. What makes it genuinely useful is the permission layer — you decide which mutations an assistant can touch, so it can manage migrations and update your cloud config without free rein over production. Compared to generic database MCP servers, this one speaks your schema natively and knows your project's deployment context, which cuts a lot of hallucinated SQL. It's aimed at teams already on the Nhost stack who want AI pair-programming that respects guardrails. If you're building full-stack apps and tired of copy-pasting schema into chat windows, this is worth a serious look.
📇 🏠 🍎 🪟 🐧 - Draw and edit draw.io diagrams from your project's files on a live browser canvas you can also edit by hand, with version history, undo, AWS, Azure and GCP shape libraries and PNG or SVG export
TypeScript · 36,455 stars
View on GitHub →What we said about next-ai-draw-io
Picture describing an architecture in plain English and watching the boxes and arrows assemble themselves on a live draw.io canvas you can still grab and rearrange by hand. That's what next-ai-draw-io does, and the detail I appreciate is that it treats the AI as a collaborator rather than a replacement for your own editing. You get version history, undo, and the full cloud shape libraries for AWS, Azure, and GCP, so the output actually looks like the diagrams your team already reviews. Export lands as clean PNG or SVG, which slots neatly into docs or slides. Compared to pure text-to-diagram tools like Mermaid, where you're locked into syntax, this keeps the familiar visual editor in the loop, so you're never fighting the layout engine. It's built in TypeScript as a web app, which makes it approachable to fork or self-host. This is for engineers who hate the blank-canvas moment and would rather refine a rough first draft than draw every node themselves. A genuinely practical use of generative models.
🐍 🏠 🍎 🪟 🐧 - Identify and remove AI watermarks and provenance metadata (Gemini and Veo marks, SynthID, C2PA, EXIF) in images and video you generated yourself, plus provenance metadata in audio. Runs as uvx remove-ai-watermarks mcp
Python · 6,161 stars
View on GitHub →What we said about remove-ai-watermarks
This tool strips out the invisible and visible markers that generative models bake into the content you create—think Gemini and Veo watermarks, SynthID, C2PA signatures, and EXIF provenance tags across images, video, and audio. It's called remove-ai-watermarks, and it runs cleanly through uvx, even exposing itself as an MCP server so you can wire it into an agent workflow.
Now, let's be honest about what this is: a provenance scrubber, and that cuts both ways. For developers building pipelines who need clean, metadata-free assets they generated themselves, it saves a lot of manual fiddling with exiftool and custom scripts. But stripping SynthID and C2PA also removes the very signals meant to tell people something was AI-made, so think carefully about where your outputs end up.
Compared to scattered command-line one-offs, the value here is consolidation—one interface for multiple watermark types and formats. The MCP integration is the genuinely interesting part, letting automated agents handle this as a step. Six thousand stars says plenty of people want it; just use it with your eyes open.
🐍 ☁️ 🍎 🪟 🐧 - Screeners, multi-timeframe technical analysis, smart-money and volume scans, and walk-forward backtests for crypto, stocks, FX and futures
Python · 5,159 stars
View on GitHub →What we said about tradingview-mcp
This one hands your AI assistant direct access to TradingView-style market analysis — screeners, multi-timeframe technicals, smart-money and volume scans, plus walk-forward backtests across crypto, stocks, FX, and futures. It's called tradingview-mcp, and the key word there is MCP: it's a Model Context Protocol server, which means instead of writing a script that calls an API, you let a model like Claude request these scans conversationally and reason over the results.
What makes that interesting is the walk-forward backtesting. A lot of trading toolkits give you a single clean backtest that quietly overfits; walk-forward testing re-optimizes on rolling windows, which is a far more honest read on whether a strategy actually survives changing conditions.
Who's this for? Quant-curious developers and traders who already live inside an LLM workflow and want market data as a first-class tool rather than a separate dashboard.
Compared to raw TradingView scraping or a bare charting library, the value here is the glue — standardized, AI-callable, and cross-asset. Just remember: a clean scan is a signal, not a trade.