🛠️ 135 Agents Your Claude Setup Is Missing

Two repos dropped this week that I keep coming back to.

The first is dev-browser: a Claude Skill that hands your agent a real web browser. Not a headless scraper. A live browser the agent opens, navigates, and reads from, using Claude's native skill protocol. 6,504 stars already. The pattern it unlocks, agents that can look things up mid-task without you hardcoding every source, is the one I've been wiring by hand for six months. Someone just packaged it cleanly.

The second is Uber's ADR. Uber runs enterprise AI agents at scale and they open-sourced the security layer: observability hooks, threat detection, benchmarking for adversarial inputs. 629 stars and three days old. The interesting thing here is not the code, it's the signal. When a company Uber's size ships their agent-defense stack publicly, the threat model has crossed from "theoretical risk" to "production incident we already had." That's the tell.

Elsewhere: options intelligence for agents, a Claude Code toolkit with 135 pre-built agents, and an MCP server that gives any agent full desktop control. Busy Wednesday.

New here? This part's for you. (Operators, skip to The Drops.)

What is the terminal?

The terminal is a text-only window where you type instructions directly to your computer. No buttons, no menus, just you typing a command and the computer responding with text. Think of it like texting your computer instead of pointing and clicking. It sounds intimidating but most AI tools use it for exactly one job: starting something up. After that, you switch back to your normal browser or app.

Do this today

Open your AI coding tool (Claude, ChatGPT, anything you use) and look for a setting called "Skills," "Plugins," or "Extensions." You do not need to install anything. Just note what is already connected. That list is your agent's current reach.

How to use today's picks

Today's Drops include 6 repos: browser skills that let AI agents navigate the web, security layers for enterprise deployments, options intelligence tools, and a 135-agent Claude Code toolkit. For each one: read the one-line pitch, click through, and skim the README (the project's front-page notes explaining what it does and how to get started). You are just browsing. Install only when something solves a real problem you are already hitting.

Plain English

- Skill: a packaged instruction set you load into an AI agent so it can do a specific job, like using a browser or analyzing color, without you writing the code yourself. - Stars: the GitHub equivalent of a thumbs-up from another developer. More stars usually means more people have tried it and found it useful.

We read your replies

You told us you are here to learn, so we built this section for you. Hit reply with any word from today's issue that made you pause. We will decode it in the next one.

[Affiliate] Browse AI, a browser skill lets your agent read a page live, but it does not solve pulling the same structured data off that page every day. Browse AI turns any site into an API without code, ships 250+ prebuilt robots, and re-runs itself when the page changes. Your first website extraction is free. We may earn a commission.

[Skill] dev-browser, gives your Claude agent a real web browser as a native skill. 6,504 stars. Your agent can navigate, read pages, and extract live information mid-task without you wiring a separate scraper.

[Repo] uber/ADR, Uber's open-source agent defense layer: observability hooks, threat detection, and security benchmarking for AI agents in production. 629 stars. The fact that Uber is shipping this publicly tells you the threat model is already real at scale.

[Skill] AlphaGBM/skills, 29 real-data options intelligence skills for Claude Code, Cursor, and other coding agents. 1,691 stars. If you are building anything in finance or trading, this drops live market data straight into your agent's context.

[Repo] rohitg00/awesome-claude-code-toolkit, 135 agents, 35 skills, 42 commands, 176+ plugins, 20 hooks, 14 MCP configs, and 26 companion apps for Claude Code, all in one curated list. 2,440 stars. This is the fastest way to audit what your current Claude Code setup is missing.

[Repo] ai-for-developers/awesome-ai-coding-tools, a curated list of AI-powered coding tools across categories. 1,948 stars. Useful when you need to compare the full tool landscape before committing to a stack.

[Skill] meodai/skill.color-expert, an agent skill covering color science: color spaces, APCA/WCAG accessibility, palette generation, and historical color theory. 540 stars. Worth loading if you are shipping UI work and want your agent to stop recommending inaccessible contrast ratios.

[Repo] microsoft/graphrag, Microsoft's modular graph-based RAG system. 35,244 stars. When flat vector search is failing on complex, relationship-heavy documents, this is the architecture you reach for.

[Repo] wuyoscar/GPT-Image2-Skill, a prompt gallery, image library, agentic skill, and CLI for OpenAI image generation and editing. 4,150 stars. Handy if you need a repeatable image-generation workflow inside an agent pipeline rather than a one-off web UI.

[Repo] AMAP-ML/LongHorizon-Harness, a long-horizon computer-use harness for running AI agents across desktop apps and the CLI over extended periods while preserving task state. 167 stars. Early, but the state-preservation piece is the part most agent frameworks quietly skip.

[MCP] nuphus-mcp, a desktop automation MCP server that gives any AI agent full computer control: screen, windows, mouse, keyboard, and Chrome, via the Model Context Protocol over stdio. The non-obvious setup detail is that it runs locally over stdio so there is no cloud relay, your keystrokes and screen data stay on your machine. Wire it when you need an agent to operate software that has no API.

- ChatGPT Work's architecture, reconstructed. Latent Space published an external teardown of how ChatGPT Work handles memory, proactivity, scheduling, browser use, plugins, and tools for its mass-market user base. The pattern that stands out for operators: proactive scheduling and state-held memory are table stakes now, not differentiators. If your agent is stateless, you are already behind the consumer baseline. (Latent Space)

- MiniMax-H3 runs locally on Apple Silicon. Simon Willison flagged that MiniMax-H3, a general-purpose omni-modal model that accepts text, images, and audio, now has an MLX port for local Apple Silicon deployment. Open weights, runs offline. If you need a local multimodal model that does not require NVIDIA hardware, this is worth a look. (Simon Willison)

- Vercel deploys are 33% faster for ISR-heavy apps. If you are shipping Next.js apps with large numbers of pre-rendered pages, deploy times dropped up to 33% with no config change required. Ships automatically. (Vercel)

- LLMs reward expertise, not just prompting. Sean Goedecke argues that the delta between a weak and strong LLM output has less to do with prompt structure and more to do with how much domain knowledge the operator brings to the conversation. Worth reading if you are still chasing prompt templates as a primary lever. (seangoedecke.com)

The move: pre-load your agent's context with a skills audit before every session.

The problem: most operators start a Claude Code session from scratch every time, rediscovering the same tools, writing the same setup prompts. The fix costs three minutes.

1. Open your CLAUDE.md (or equivalent project instructions file). Add a section called ## Available Skills and list every skill or plugin you have installed, one line each, with a one-sentence description of what it does. 2. At the top of any new session, tell the agent: "Check Available Skills before proposing a solution." One line. 3. When you add a new skill (today's dev-browser, for example), append it to the list immediately. Treat it like a changelog entry.

You will know it worked when the agent stops suggesting you manually scrape a URL after you installed dev-browser, because it already knows the browser skill is there.

The real payoff is cumulative: every session starts with a current, accurate map of what your agent can actually do.

Five green checks in one working session, most of them mine, and not one of them could have failed.

A check that cannot fail is not evidence. It is decoration that looks like diligence.

In a single working session I hit the same shape five separate times, and most of them were mine:

• A search across our own files came back clean, because the pattern I searched for could never have matched anything.
• A counter I read as proof that nothing had gone wrong turned out to be a dead instrument, unwired for weeks. It reads zero forever.
• A file-restore step reported success while doing nothing at all.
• A rule written to catch one specific bad case flagged a file already fixed, and would have missed the case I actually cared about.
• A status check invented a value for a document that never had one, because I looked in the whole file instead of the one section where that field means anything.

Every one returned exactly what I was hoping to see. That is the tell, and it is uncomfortable, because the hope arrives before the result and quietly lowers the bar for it.

The rule I landed on: a rule that names one instance has silently claimed a whole class. Write "this block type breaks" and you have implied every other one is fine, without checking a single one.

The countermeasure is cheap enough that there is no excuse. Before you trust a check, prove it can go both ways. Break the thing on purpose and confirm the check screams. Fix it and confirm it goes quiet. Two extra commands. And say where you ran it from, because with all five of these, the vantage point was the entire defect.

What is the last green result you got that you never proved could turn red? Hit reply. We read every one.

Nine drops, one desktop-control MCP, and the clearest sign yet that agent security is moving from a side conversation to a production requirement. The browser skill and the Uber defense layer are the two I would actually open today.

Before you go: we started a room for people actually building with agents. Tell us what you want us to build next, ask when you get stuck, and show the setup you are proudest of. Join the community →

See you Thursday.