🛠️ Instatic Killed Webflow. It's Free.

Monday. Let's get into it.

The biggest story in this issue is not a model release. It's a repo. Instatic is a self-hosted, agentic visual CMS that outputs clean static pages and already has 5,558 stars. It has users, roles, plugins, content management, and a database. It runs on your own server. The thing that would have cost your client $20,000 in Webflow development and a monthly subscription is now a clone-and-deploy away.

That's the week's sharpest data point: the price floor for shipping a real product just dropped again.

Alongside that: Simon Willison surfaced an investigation into the relay market, a shadow economy built on reselling LLM tokens at a discount using pooled API keys. Fraud infrastructure, running in the open. That's a threat to your cost stack and your security posture, and I'll cover it in Signals.

Claude Opus 5 also landed last week at roughly half the price of Fable-level performance. That's worth a line but it's not the lead, the operator consequence is cost, and the repo drops are the action.

Eight repos and a skill below. One MCP. A handful of signals that shift something real.

New here? This block is for you. If you've been building with AI for a while, the next section has your repos.

What is open source?

When a piece of software is "open source," the people who built it have published the code for anyone to read, copy, and modify. Think of it like a recipe posted publicly online: anyone can use it, improve it, and share their version. Closed software is more like a restaurant dish, you can order it, but the chef never tells you what's in it. Most of the repos in today's issue are open source, which means you can run them yourself, inspect what they do, and change them to fit your needs. There's no vendor to call, and no monthly bill if you host it yourself.

Do this today

Find one piece of software you pay for every month (a website builder, a writing tool, a scheduling app). Search that name on GitHub with the word "alternative." You'll often find an open-source version with thousands of stars. Don't install anything yet, just look at what exists. That's the whole exercise.

How to use today's picks

Today's issue has 8 repos (those are GitHub projects you can copy and run), plus a Claude Code skill and an MCP server. For each one, I've written a one-line description of what it does. Click through if it sounds relevant. When you land on the page, scroll past the top image and read the README (the project's front-page notes explaining what it does and how to start). You don't need to install anything today. You're building a map of what's possible.

Plain English

Repo: Short for "repository." It's a folder of code hosted on a site like GitHub, the way a Google Drive folder holds your documents. When we say "clone a repo," we mean copy that folder to your own computer.

Stars: On GitHub, users "star" a repo the way you'd bookmark a page. High star counts signal real people found it useful, not just that it exists.

We read your replies

You told us you're here to learn, so we built this block for you. Every issue, one concept in plain terms. If there's a word you keep seeing in AI content that you can't quite place, reply with it and we'll decode it here.

[Repo] Instatic, An open-source, self-hosted visual CMS with agentic output, clean static pages, users, roles, plugins, and a database baked in. Webflow and Framer charge you monthly for the same feature set. This one has 5,558 stars and runs on your own server. If you build client sites, this changes the math.

[Repo] Chat2DB, An AI-driven SQL client and database GUI supporting MySQL, PostgreSQL, Oracle, SQLite, ClickHouse, and more. 27,037 stars. You talk to your database in plain language; it writes and runs the query. If you are still writing raw SQL for every report, you are doing it the slow way.

[Repo] drona23/claude-token-efficient, One CLAUDE.md file that keeps Claude responses terse and reduces output verbosity on heavy workflows. Drop it in, no code changes. 5,876 stars. Every token you save on a long agentic run is money you keep; this is the cheapest optimization you will make this week.

[Repo] kbwo/ccmanager, A session manager for coding agent CLIs: Claude Code, Gemini CLI, Codex CLI, Cursor, Copilot, Cline, and others. 1,202 stars. If you are running more than one agent at a time, you already know the pain of tracking which session is doing what. This is the fix.

[Repo] qdrant/qdrant, High-performance, cloud-native vector database built for AI-scale search. 33,595 stars. The anchor vector store for anyone building retrieval-augmented anything. If you are not running your own vector DB and you are doing RAG, you are paying someone else's margin.

[Repo] huggingface/smolagents, A barebones library for agents that think in code, from Hugging Face. 28,546 stars. No framework overhead, no abstraction layers hiding what the agent actually does. Good for operators who want to understand what is running, not just that something is running.

[Repo] assafelovic/gpt-researcher, An autonomous agent that conducts deep research on any topic using any LLM provider. 28,649 stars. Point it at a question; it works through sources and hands you a structured report. The research leg of most client workflows, automated.

[Repo] TheDecipherist/claude-code-mastery, A complete guide to Claude Code: CLAUDE.md files, hooks, skills, MCP servers, and slash commands. 541 stars and growing. If you are new to Claude Code or you are using it without hooks and skills, this is the fastest path to running it properly.

[Skill] NulightJens/humanizer-stack, A two-pass pipeline that removes AI writing tells from outward-facing text: a surface pass plus a structural pass grounded in the StoryScope study. Packaged as Claude Code Skills. 91 stars. If you are shipping AI-drafted content to clients or to an audience, this is the last step before it sounds like a person wrote it.

[Affiliate] Runpod, Today's theme is owning your stack instead of renting it, and that argument runs out at the GPU unless you have somewhere to put the open models. Runpod rents GPU compute on demand, so you can self-host the open-weight models in this issue without buying hardware. Worth a look if the relay-market story made you want your own endpoint. We may earn a commission.

[MCP] mixelpixx/Konnect, AI-assisted PCB design for KiCAD 10. A single Rust binary that exposes 171 schematic, layout, routing, design-review, and manufacturing tools to Claude or any LLM you choose. If you build hardware products or work with clients who do, this is the first credible AI layer for the KiCAD workflow. Native plugin, no wrapper gymnastics.

Three things that shift the ground this week:

Chinese open models are inside U.S. infrastructure. Kimi K3, DeepSeek, and related Chinese models are gaining real U.S. users, including Mozilla's CTO. Cheaper, open-weight, and increasingly capable. The operator question is not whether they are good enough; it is whether your clients will ask you to run them and what the compliance answer is. (AP News)

The relay token market is running fraud at scale. Simon Willison flagged an investigation by Matt Lenhard into a shadow market built on pooling API keys to resell LLM tokens at a discount. These relay networks are active, commercial, and growing. If your agent stack uses any third-party API proxy, you may be downstream of this. Check your provider's terms and confirm you are hitting the actual vendor endpoint. (Simon Willison)

Claude Opus 5 ships at roughly half the Fable price. Anthropic released Claude Opus 5 with self-reported Fable-level performance at a significantly lower price point. The operator case is straightforward: same capability tier, lower per-token cost. If you are running Fable on heavy workflows, the switch math is worth running this week. (Anthropic)

The move: Drop the token governor into any Claude Code project.

Here's what I'd do with drona23/claude-token-efficient today:

1. Clone the repo and copy the CLAUDE.md file into the root of your active Claude Code project.

2. Open it and scan the directives. You will see instructions that cap response length, suppress preamble, and kill the "let me explain what I'm about to do" paragraphs Claude defaults to on long tasks.

3. Run your most verbose existing workflow once with it in place. Compare the token count before and after.

4. Adjust the verbosity cap to your tolerance: tighter for bulk pipeline runs, looser for exploratory sessions.

You will know it worked when your agent stops narrating its own reasoning on every tool call and just does the work. On a multi-step agentic run, the cost difference is not marginal.

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. This week: a dashboard that showed us a perfectly real figure for a business that was not ours.

A dashboard will render a confident number for whatever account it happens to be pointed at, and it looks identical when it is pointed at the wrong one.

We went to read a live metric for the newsletter. The panel loaded, the figure was clean, the charts were full. It was also the wrong business. The tool had defaulted to a different account than the one we meant, and every number on the screen belonged to something else. Nothing on the page announced this. A real value for the wrong entity is indistinguishable from a real value for the right one, right up until you check the small print that says whose data you are looking at.

The trap is that an empty or broken panel warns you. A fully populated panel for the wrong account reassures you, in exactly the tone you were hoping for, and sends you off to make decisions on a stranger's numbers.

So we added one step before reading any figure: confirm the account id, the dataset, the property, whatever names the owner of the data, before we so much as glance at the value. The identity of the source is part of the reading. A number without its account attached is not a number yet.

The habit we kept: read whose it is before you read what it is.

What decision did you almost make on the wrong account's data? Hit reply. We read every one.

The relay token story is the one I keep thinking about. A shadow economy that profits from your API spend exists, runs quietly, and most operators have no idea they may be inside it. Check your proxy chain this week. That's the one action that matters most from this issue.

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

See you Tuesday.