Weekly · AI agent toolkit · week 38, 2026
andrej-karpathy-skills + 7 more must-have coding-agent tools · week 38, 2026
This week's must-have repos for AI coding-agent users — andrej-karpathy-skills, AutoGPT, skillmem, 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
213,286 stars
View on GitHub →What we said about andrej-karpathy-skills
Here's one that's less code, more philosophy. It's a single CLAUDE.md file that distills four behavioral guidelines for working with Claude Code, drawn from Andrej Karpathy's public notes on where LLM-assisted coding tends to go sideways. No install, no dependencies — you just drop it into your project root and your agent picks it up.
What I like here is the humility of the format. Instead of another sprawling framework, it's a set of guardrails: nudging the model to slow down, verify before it acts, and avoid the confident-but-wrong patterns we've all been burned by. Think of it as a shared code review culture, but for your AI pair.
Compared to prompt-engineering repos or heavyweight agent toolkits, this wins on friction — you can adopt it in thirty seconds and actually read the whole thing. It's for anyone already living in Claude Code who wants better defaults without rearchitecting their workflow.
Whether four rules justify the star count is your call, but the idea is sound.
If you want more finds like this, the newsletter link's in the description.
by dyoshikawa - A Node.js CLI tool that automatically generates configs (rules, ignore files, MCP servers, commands, and subagents) for various AI coding agents. Rulesync can convert configs between Claude Code and other AI agents in both directions
TypeScript · 1,426 stars
View on GitHub →What we said about rulesync
So here's a problem you probably didn't know you had until you hit it: every AI coding agent wants its own config format. Claude Code has its rules, Cursor has its own, and each one expects you to hand-maintain ignore files, MCP server definitions, commands, subagents — the whole stack, times however many tools your team uses. Rulesync is a TypeScript CLI that treats one source of truth and generates the rest, converting configs between Claude Code and other agents in both directions.
What I like here is the bidirectional part. Most sync tools are one-way exporters; this actually reads existing setups back in, so you can migrate without starting from scratch. If you're on a team where half the devs use one agent and half use another, this is the glue that keeps everyone on the same rules.
It's early, and format coverage will make or break it, but the instinct is right — config shouldn't lock you into a vendor.
If you want more finds like this each week, the newsletter link is in the description.
Works With Any Agent
Facebook
Python · 187,377 stars
View on GitHub →What we said about AutoGPT
AutoGPT is one of those projects that basically kicked off the whole autonomous agent conversation back in early 2023. The idea was simple but wild at the time: give a language model a goal, and let it break that goal into tasks, execute them, and loop until it's done — no human babysitting each step. What's interesting is how the project has actually evolved since that viral moment. It's shifted from a scrappy command-line experiment into a more structured platform for building and deploying agents, with a visual builder and reusable blocks. That matters because the honest lesson from those early days was that fully hands-off agents drift, hallucinate, and burn tokens fast. So the newer direction leans toward giving you guardrails and composability instead of pure autonomy. If you're weighing it against something like LangChain or CrewAI, AutoGPT is aiming more at the low-code, deployable-workflow crowd than the library-first developer. Worth a look if you're prototyping agent ideas without wanting to wire everything from scratch. For more breakdowns like this, the newsletter link is in the description.
🐍 🏠 🍎 🪟 🐧 - Self-improving skill memory for coding agents: skills that help get reinforced, unused ones fade on an Ebbinghaus decay curve. $0 write path (SQLite FTS5 + local ONNX embeddings, no LLM calls), bilingual EN/RU hybrid search, SHA256 tamper-evident history, deterministic LongMemEval benchmark in-repo (hit@5 0.871). 8 mem_ tools; deep Claude Code hook integration optional. pip install skillmem
Python · 4 stars
View on GitHub →What we said about skillmem
Here's something clever from liza-studio. Skillmem tackles a real problem with coding agents: they relearn the same things over and over because nothing persists between sessions. This gives your agent a memory that behaves a bit like human memory — skills that actually get used get reinforced, and unused ones fade on an Ebbinghaus decay curve. So instead of an ever-growing pile of stale context, you get a memory that prunes itself.
What I appreciate is the engineering restraint. The write path costs you nothing — SQLite full-text search plus local ONNX embeddings, no LLM calls burning tokens every time you save. That's a genuinely different tradeoff from vector-DB-heavy memory frameworks that charge you on every write. Bilingual English-Russian search and SHA256 tamper-evident history are thoughtful touches, and they ship a deterministic benchmark right in the repo, which is rare and honestly builds trust.
It's early, four stars, but the ideas are worth watching if you're building agent tooling. If you want more finds like this before they blow up, the newsletter link's in the description.
🐍 🏠 - The policy enforcement plane for MCP -- deterministic admission and egress policy, attributable audit, and SIEM export for your MCP server fleet. MIT, self-hosted, no SaaS
Python · 15 stars
View on GitHub →What we said about mcp-hangar
Here's a repo that's early but pointing at a real problem. As teams wire up more and more MCP servers, you end up with this sprawl of tools that can read files, hit APIs, touch internal systems—and almost no one has a clean answer for what those servers are actually allowed to do. That's the gap mcp-hangar is going after. Think of it as a policy layer that sits in front of your MCP fleet: deterministic rules for what gets admitted, what data can leave, plus an audit trail you can actually attribute to a specific caller, and SIEM export so your security folks aren't flying blind.
What I appreciate here is the posture—MIT licensed, self-hosted, explicitly no SaaS. If you're in a regulated shop or you just don't want your tool traffic routed through someone else's cloud, that matters. It's Python, it's fifteen stars, so temper expectations—this is ground-floor. But it's the kind of infrastructure that becomes obvious in hindsight.
If governance-for-agents is on your radar, we go deeper in the newsletter—link's in the description.
🐍 🏠 🍎 🪟 🐧 - Local-first memory for AI agents about the people in your life: explainable identity resolution, relationships, roles, facts, interactions, reminders, and communication guidance in one SQLite file you own. Imports are review-gated, sensitive disclosure is operator-gated, and ordinary operation makes no network calls. uvx --from people-context people-context
Python · 30 stars
View on GitHub →What we said about people-context
Here's a small project doing something genuinely thoughtful. people-context gives your AI agents a memory of the actual humans in your life — who they are, how they relate, roles, facts, past interactions, even communication guidance — and it lives entirely in one SQLite file that you own. No cloud, no account, no telemetry. Ordinary operation makes zero network calls.
What I like here is the discipline around trust. Imports are review-gated, so nothing lands in your memory without you seeing it first, and disclosure of sensitive details is operator-gated. That's a real answer to the "my assistant knows too much and shares it too freely" problem that most agent memory frameworks just hand-wave past.
Compared to vector-store memory that blurs everything into embeddings, this leans explainable — identity resolution you can actually audit. It's early, thirty stars, but it's for developers building personal agents who care about privacy and provenance, not just recall.
If local-first tooling is your thing, the newsletter rounds up more like this every week — link's in the description.
🐍 ☁️ 🏠 🍎 🪟 🐧 - Cross-agent failure intelligence. Before an agent retries a failed tool call, check_tool_failure reports whether other agents are hitting the same failure right now and which recovery actually worked, with a Wilson-score confidence. Privacy-by-schema: no prompts, keys, or payloads stored. Hosted remote server at https://failecho.com/mcp, or self-host
Python · 2 stars
View on GitHub →What we said about failecho
Here's a clever little idea that punches above its two stars. FailEcho tackles something most agent frameworks quietly ignore: when a tool call blows up, your agent retries blind, with no idea whether that failure is systemic or just bad luck. What FailEcho does is pool failure signals across agents, so before you retry, `check_tool_failure` tells you if others are hitting the same wall right now — and crucially, which recovery actually worked, weighted by a Wilson-score confidence interval so you're not chasing one lucky fluke.
What I appreciate is the privacy-by-schema design. It stores no prompts, no keys, no payloads — just failure fingerprints. That's the kind of decision that makes shared telemetry actually adoptable in a real org.
Compared to rolling your own retry logic or leaning on generic observability tools, this is purpose-built for the agent-to-agent feedback loop. You can hit their hosted MCP endpoint or self-host if you'd rather keep it in-house.
It's early, but the concept is sound. If you like catching these before they blow up, the newsletter link's in the description.
🐍 🏠 🍎 🪟 🐧 - Persistent agent memory with a first-class correction channel: supersede or revert a fact so recall stops returning the stale value, an echo guard that blocks a restated old value from resurrecting it, and receipted erasure (a signed, content-free tombstone that makes a deletion provable). No LLM on the write path, zero dependencies, the core in one file, 73 MCP tools. uvx --from "inspeximus[mcp]" inspeximus-mcp
Python · 6 stars
View on GitHub →What we said about inspeximus
Let's talk about inspeximus, which tackles a problem most agent memory systems quietly ignore: what happens when a stored fact is just wrong. Plenty of tools let you write memories. Very few give you a clean way to correct them. Here you can supersede or revert a fact so recall stops surfacing the stale value, and there's an echo guard that stops an old value from sneaking back in when your model restates it. My favorite detail is receipted erasure — a signed, content-free tombstone that makes a deletion actually provable, which matters if you're ever thinking about audits or compliance.
The design choices are opinionated in a good way: no LLM on the write path, zero dependencies, core in a single file, and 73 MCP tools ready to go. That determinism is the real pitch versus vector-store-heavy approaches that blur what's remembered and why.
It's early — six stars — so kick the tires before production. If you want more finds like this, the newsletter link's in the description.