Weekly · AI agent toolkit · week 45, 2026
skills + 7 more must-have coding-agent tools · week 45, 2026
This week's must-have repos for AI coding-agent users — skills, claude-code, Aelios, and more: Claude Code-first, plus tools that work with any agent, drawn out of the noise.
Claude Code Tool
by Anthropic - Anthropic's official repository for Agent Skills — the SKILL.md format, a skill template, and example skills, the same format Claude Code loads natively
Python · 187,209 stars
View on GitHub →What we said about skills
This gives you a clean, repeatable way to teach an agent new capabilities by writing them down in a structured Markdown file instead of hard-coding logic or cramming everything into one giant system prompt. It's Anthropic's skills repo, and it standardizes the SKILL.md format that Claude Code already loads natively, plus a template and working examples to copy from.
What makes this worth a look is the separation of concerns. A skill becomes a portable, version-controlled unit — documentation, instructions, and supporting files bundled together — so your agent pulls in context only when it's actually relevant. That keeps token usage lean and behavior predictable.
If you've been wrangling sprawling prompts or juggling MCP servers for everything, this is a lighter-weight middle ground. It's aimed at developers building on Claude who want reusable, shareable agent behavior rather than one-off scripting.
The caveat: it's tied to Anthropic's ecosystem, so it's less of a universal standard and more of a strong convention. Still, the format is simple enough that the ideas travel well.
A delectable showcase of top tier skills, ambidextrous agents, scintillating status lines, top notch developer tooling, and also we have plugins. Suitable for beginners and veterans, with an emphasis on code quality, security, and originality._
TypeScript · 154,462 stars
View on GitHub →What we said about claude-code
This one puts an AI coding agent directly in your terminal, letting it read your codebase, edit files, run commands, and work through tasks the way you'd actually walk a junior dev through them. It's called claude-code, from Anthropic, and the repo you're looking at is really the ecosystem around it — skills, plugins, status lines, and tooling meant to extend what the agent can do.
What stands out to me is the shape of this. Instead of another chat box bolted onto an editor, it leans into the command line, which means it slots into whatever workflow you already have — vim, tmux, your own scripts. The plugin and skills angle is the interesting part, because it's a bet that no single vendor nails every workflow, so they're handing you the hooks.
Compared to Copilot or Cursor, this trades the polished IDE experience for composability and control. If you live in the terminal and care about reviewing every change before it lands, that tradeoff probably feels like a feature, not a compromise.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond
JavaScript · 296,748 stars
View on GitHub →What we said about ECC
This one sits between you and your coding agent, shaping how it remembers context, follows learned patterns, and stays inside safe boundaries while it works. It's called ECC, and the pitch is that your AI assistant—whether that's Claude Code, Codex, Cursor, or Opencode—shouldn't start every session as a blank slate. ECC layers in skills, instincts, and persistent memory so the agent carries forward what it's learned about your codebase instead of relearning it every prompt.
What makes this interesting is the harness framing. Most tooling treats the model as the whole system, but ECC treats the scaffolding around it as the real lever—security rails, a research-first workflow, reusable behaviors. That's closer to how experienced engineers actually think about reliability.
If you're bouncing between agents and tired of each one having its own quirks and no shared memory, this aims to be the connective tissue. It's for people already living inside these tools daily, not newcomers. Worth a look if your agent keeps making the same mistakes you corrected yesterday.
The agent that grows with you
Python · 260,344 stars
View on GitHub →What we said about hermes-agent
This one tries to give an AI assistant something closer to persistent memory, so it actually remembers your projects, preferences, and past conversations across sessions instead of starting fresh every time. It's called hermes-agent, from the folks at Nous Research, and the framing here is "the agent that grows with you" — which in practice means a memory layer wrapped around a Python agent loop. Why does that matter? Most agent frameworks are stateless by default, and you end up re-explaining your codebase every morning. If this handles memory retrieval well, that's a real quality-of-life shift for anyone building a personal dev companion. Compared to something like LangChain or CrewAI, which lean toward orchestration and multi-agent workflows, Hermes feels more focused on the single long-lived assistant that knows you. It's aimed at developers comfortable wiring up their own tools and models, not folks wanting a plug-and-play app. Worth a look if you've been frustrated by agents with goldfish memory — just check how they handle memory decay and privacy before you trust it with everything.
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote
JavaScript · 184,004 stars
View on GitHub →What we said about ponytail
This one teaches your AI agent to push back before it writes anything, asking whether the feature even needs to exist. It's called ponytail, and the philosophy is deliberately contrarian — instead of generating code on demand, it channels that senior engineer who answers your pull request with "do we actually need this?" In practice, it wraps your agent in a planning layer that favors deleting code, reusing existing utilities, or just closing the ticket over shipping new abstractions. Why does that matter? Most coding assistants are eager to please, which means they happily add complexity you'll maintain forever. Ponytail inverts the incentive. It's aimed at teams drowning in agent-generated sprawl, and it pairs nicely with tools like Aider or Cursor rather than replacing them — it's the restraint those tools lack. The tradeoff is real: sometimes you do want the code written, and arguing with a skeptical agent gets tiring. But for anyone who's inherited a codebase bloated by good intentions, that friction is the whole point.
A cross-platform desktop All-in-One assistant for Claude Code, Codex, OpenCode,DeepSeek Harness and more. Only official website: ccswitch.io
Rust · 151,684 stars
View on GitHub →What we said about cc-switch
Imagine juggling four or five different AI coding agents, each with its own config files, API keys, and environment setup scattered across your machine. This is cc-switch, a cross-platform desktop app that acts as a single control panel for Claude Code, Codex, OpenCode, DeepSeek Harness, and a growing list of others. What makes it click is that it's built in Rust, so you get a lightweight native binary rather than yet another heavy Electron wrapper eating your RAM.
The real value here is the switching workflow. Instead of manually editing dotfiles when you jump between agents or swap providers, you flip between profiles from one interface. For developers who evaluate tools constantly, or teams standardizing on multiple models, that's a genuine friction-killer.
Compared to rolling your own shell scripts or living inside each tool's separate CLI, cc-switch centralizes the mess. If you're already committed to a single agent, you may not need it. But if your terminal is a zoo of coding assistants, this is worth a look.
Works With Any Agent
📇 ☁️ - Self-hosted long-term memory gateway for AI clients: runs on your own Cloudflare account (Workers + D1 + Vectorize), captures conversations, consolidates them overnight, and recalls them across Chatbox, Claude Code, and Codex via API or MCP tools
TypeScript · 142 stars
View on GitHub →What we said about Aelios
Picture a memory layer that sits between you and your AI tools, quietly storing every conversation and surfacing the relevant bits later, all running on infrastructure you actually control. That's Aelios, a self-hosted memory gateway built entirely on your own Cloudflare account — Workers for the logic, D1 for storage, and Vectorize for semantic recall. What I find clever here is the overnight consolidation step. Instead of dumping raw transcripts into a vector store and hoping retrieval works, it periodically compresses and reorganizes what it's learned, which is closer to how human memory actually settles. It speaks to Claude Code, Codex, and Chatbox through both a plain API and MCP tools, so you're not locked into one client. Compared to hosted memory services like Mem0 or Zep, the trade-off is clear: you handle the setup, but your conversation history never leaves your tenant. This is for developers who want persistent context across tools without renting someone else's database. At 142 stars it's early, but the architecture is thoughtful and the Cloudflare free tier makes it genuinely cheap to try.
🐍 ☁️ 🏠 - Email inboxes for AI agents via Sendmux: receive, search and send mail
PHP · 62 stars
View on GitHub →What we said about qubrics-sdk
Here's a clever one: this gives your AI agents an actual email inbox, so they can receive messages, search through past mail, and send replies on their own. It's called qubrics-sdk, and it leans on a service named Sendmux to handle the heavy lifting of mail delivery and parsing. What makes this interesting is the shift in thinking — instead of bolting email onto an agent with brittle IMAP scripts, you get a proper interface built for programmatic use. Think support bots that triage incoming tickets, or a research agent that subscribes to newsletters and summarizes them for you. The curious part is the language: it's written in PHP, which you don't often see in agent tooling, so if your stack already lives in Laravel or a legacy PHP app, this slots in naturally without dragging a Python runtime alongside it. At sixty-two stars it's early, and you'll want to check how much depends on Sendmux versus what you control. But the idea of email as a first-class agent channel is worth watching.