🛠️ Garry Tan's 23-Tool Claude Setup

Garry Tan published his exact Claude Code setup this week and then pushed it to GitHub. Twenty-three opinionated tools. Each one playing a specific role: CEO, Designer, Eng Manager, Release Manager, Doc Engineer, QA. Not a demo, not a concept deck. The actual config he runs. It has 125,302 stars.

That number matters less than the signal: the people building the most are publishing their whole operating layer now. Not just the tool, the judgment that wires it together.

I also found an open-source drop of Claude Cowork (call it openwork), a trading agent framework that somehow cracked 95K community stars, a full AI SRE toolkit, and a local Claude Code cost dashboard that shows you the full burn picture your Pro or Max plan hides from you. That last one is useful in a way that stings a little.

Today's issue: 8 drops worth cloning, a Stack pick for semantic code navigation, signals on GPT-5.6 pricing and the Nvidia alliance fracture, and a Playbook move on agentic research sourcing. Let's work.

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

What is inference?

When an AI model gives you an answer, it is doing inference. You send it text (your question, your prompt), and it computes a response word by word, based on everything it learned during training. Inference is the "running" step, as opposed to the "learning" step. Training happens once, on massive hardware, over weeks. Inference happens every time you hit send. When people talk about AI being expensive to run, they almost always mean inference costs, paying for the compute each time the model thinks.

Do this today

Open any AI tool you use and ask it a question you have been putting off answering yourself. Drafting a response, summarizing a document, planning a project. Notice how fast it replies. That speed is inference happening in real time, on hardware you do not own or pay for directly.

How to use today's picks

Today's Drops include 8 picks, seven repos and one Claude Code skill: tools for running multi-agent trading systems, building AI SRE agents, tracking your Claude token costs, and generating sprite sheets from a prompt, plus Garry Tan's complete setup. Over in The Stack you will also find a semantic code-navigation toolkit and an Emacs integration. For each one, read the one-line description, click through, and skim the README (the project's front-page notes explaining what it does and how to start). You do not need to install anything today. The goal is knowing what exists so you can move fast when you hit the right problem.

Plain English

GitHub, a website where developers share code publicly. Every repo in today's Drops is hosted there. You can browse, download, or copy any public project without an account.

CLI, stands for "command-line interface." A text-based way to control software, where you type instructions instead of clicking buttons. Many of today's repos are installed or run from the CLI.

We read your replies

You told us you are here to learn, so we built this section for you. If there is a word or concept showing up in AI conversations that is not clicking yet, reply with it. We will decode it in a future issue.

[Repo] garrytan/gstack, Garry Tan's exact Claude Code setup, open-sourced. Twenty-three opinionated tools playing named roles: CEO, Designer, Eng Manager, Release Manager, Doc Engineer, QA. You get the config and the judgment behind it. 125,302 stars.

[Repo] different-ai/openwork, the open-source alternative to Claude Cowork, built on opencode. If you want the coworking-agent pattern without a proprietary lock-in, this is the starting point. 18,588 stars and trending.

[Repo] TauricResearch/TradingAgents, a multi-agent LLM financial trading framework. Multiple specialist agents (analyst, risk, execution) coordinate across a shared state. Useful as an architecture reference even if you are not in fintech. 95,064 stars surfaced by the community this week.

[Repo] Tracer-Cloud/opensre, an open-source toolkit for building AI SRE agents. Monitors your systems, surfaces anomalies, and closes the loop without waiting for a human to open a ticket. 9,575 stars. Getting talked about across AI engineering circles this week.

[Repo] phuryn/claude-usage, a local dashboard for tracking Claude Code token usage, costs, and session history. Pro and Max subscribers get a progress bar. This gives you the actual numbers behind it. 2,092 stars. Run it once and you will immediately change how you prompt.

[Skill] rohunvora/x-research-skill, an X/Twitter research skill for Claude Code. Agentic search, thread following, deep-dives, sourced briefings. Useful when a story is breaking on X before it hits any publication. 1,212 stars.

[Repo] 0x0funky/agent-sprite-forge, an agent skill for generating 2D sprite sheets, transparent PNG frames, and animated GIFs from prompts. Niche, but if you are shipping a game or UI animation layer, this cuts the production loop significantly. 3,549 stars.

[Repo] openinterpreter/openinterpreter, a natural-language interface to run code on your machine. Connects LLMs directly to your OS: files, processes, the browser, whatever. The original local code-execution runtime, still the most permissive.

[Affiliate] MindStudio, Garry Tan's stack works because every tool has one clear job. If you want that shape without wiring it yourself, MindStudio builds and deploys agents across 200+ models with no code, and the free tier gives you one agent and 1,000 runs a month. We may earn a commission.

[Tool] oraios/serena, a powerful MCP toolkit for coding that provides semantic retrieval and editing. It reads your codebase symbolically, not just as text, so your agent can find references, navigate call graphs, and make targeted edits without rewriting the whole file. Think of it as an IDE layer your agent actually uses. 27,222 stars.

[Tool] manzaltu/claude-code-ide.el, Claude Code IDE integration for Emacs. If you live in Emacs, this is the missing wire between your editor and Claude Code. No browser switching, no copy-paste. 1,633 stars.

- GPT-5.6 pricing drops. OpenAI cut prices on GPT-5.6 for Luna and Terra tiers, framing it as an efficiency gain for enterprise workflows. The operator consequence: if you were holding off on GPT-5.6 for cost reasons, the floor just moved. Benchmark your current stack against it now before the gap closes. (OpenAI Blog)

- Nvidia's open-source alliance launches without OpenAI or Anthropic. Nvidia is building an open-source AI coalition and neither OpenAI nor Anthropic is in it. The fracture between open and closed camps is hardening at the infrastructure level, not just the model level. Watch which toolchain vendors pick a side. (Wired)

- Ontologies are back in agentic systems. Engineers are rediscovering structured knowledge graphs as a way to keep probabilistic agents inside deterministic boundaries. If your agents are drifting out of scope or hallucinating entity relationships, this is the architecture pattern worth reading. (Latent Space)

- LinkedIn adds an "AI slop" flag. Users can now report content that appears AI-generated. Platforms are building friction into AI-generated content at the distribution layer. If you are running any automated content to LinkedIn, this changes the risk profile. (The Verge)

- Vercel passes Server-Timing headers to the client starting August 10. CDN-level change: backend metrics like database query time and cache hit rate will now reach the browser. Useful for debugging latency in any Vercel-hosted agent app. (Vercel Changelog)

Move: Wire your research agent to three source types before it writes anything

Most research agents pull from one pool: web search. The output reads like it. Here is how to fix it in a single session config:

1. Point the agent at three named source types: community discussion (Reddit, Hacker News), primary source (GitHub READMEs, arXiv abstracts), and editorial signal (newsletters, verified analyst posts). Define each as a separate retrieval step, not a single search query. 2. Require the agent to surface at least one result from each type before synthesizing. No synthesis until all three buckets have a hit. This is a hard constraint in the system prompt, not a soft instruction. 3. Tag each source in the output by type. You will immediately see which bucket your agent is leaning on. The imbalance is usually obvious on the first run.

You will know it worked when the output stops reading like a summary of a summary and starts citing things you did not already know existed. The first session that pulls a Reddit thread AND a GitHub repo AND an editorial piece on the same topic is the moment the research layer becomes an actual asset.

Connect to today's Drops: x-research-skill handles the social layer; opensre is a good test case (community coverage is rich, primary source is a real repo, editorial is thin).

Two correct systems, run in the wrong order, quietly cost us money for months. Neither one was malfunctioning.

When two systems each have a deadline, the expensive part is usually the space between them, and it belongs to nobody.

We get subscribers from a partner network, and that arrangement has a grace period: if someone leaves quickly, it costs us nothing. Separately, we run our own list hygiene, quietly removing people who never open anything, because a list full of ghosts hurts everyone who does read.

Both of those are good. I had never once put them side by side. The grace period closes at fourteen days. Our own cleanup, when I finally went and read the live configuration instead of trusting my notes, fires at about thirty. It is a wait, then a nudge, then another wait, then another nudge, and only then the removal, and those steps add up to more than double what I thought.

So for every subscriber who was never going to open anything, we let the free window close, paid, and then removed them two weeks later. Two correct things, done in the wrong order, and no dashboard shows that, because neither system is malfunctioning. The cost lives in the seam.

The part that stings is the smaller one. My own written note said the cleanup ran at two weeks. Nobody had ever measured it. It had been true once, or it had been a guess that hardened into a fact by repetition, and it sat there for months being quoted in decisions.

What I took from it: when a vendor's clock and your clock both touch the same object, draw the two timelines on one line and look at the overlap. And any number in your own documentation that nobody has measured recently is not a fact. It is a rumour you are paying interest on.

Where do somebody else's deadline and your own overlap? Hit reply and tell us who pays for the gap. We read every one.

Ten picks today, most of them with the kind of star counts that only happen when something solves a real problem for a lot of people at once. The open-source pace is not slowing. If anything, the gap between what costs money and what you can pull from GitHub is narrowing every week. That is the whole bet here.

Before you go: we started a room for people actually building with agents. Post what you shipped this week, get unstuck, and drop what you want us to build next. Join the community →

See you Monday.