If you've used jac create --use client to scaffold a Jac full-stack project, you've seen jac-client-node and @jac-client/dev-deps in your jac.toml. They're npm meta-packages — packages that exist solely to declare a list of other packages as dependencies. The idea: one line in your config gives you React, Vite, TypeScript, and everything else you need.
Sounds clean. In practice, it's a trap. I think we should replace both meta-packages with direct dependency injection, and I want to make the case for why.
If you've ever worked with a graph database at scale, you've probably hit the N+1 query problem. It's one of those things that doesn't show up in development but absolutely kills you in production. We ran into it with Jac's tiered memory system and spent a week figuring out the right fix. This is the story of what we tried, what failed, and what actually worked.
It started in a Jaseci weekly sync. We were discussing the jac-mcp server when someone brought up the idea of an MCP playground — a place to test MCP servers the way Postman lets you test REST APIs. The idea stuck with me, but it didn't feel real until I ran into the problem myself.
I was connecting jac-mcp to GitHub Copilot, trying to understand how tools were being called and what they returned. The answer was: I had no idea. Copilot called tools based on whatever prompt and model it chose, and I was left guessing.
When I started building the jac-shadcn MCP server, I hit the same wall. Testing through an LLM client meant my results depended on the model and prompt, not the server itself. That's not testing. That's hoping.
Every browser automation script you've written has the same hidden bug: it doesn't know what it's doing. It knows the steps. It doesn't know the structure.
You write page.goto(), then page.fill(), then page.click(). Each call succeeds or fails on its own. The script has no idea that it just navigated to a login page, that the text field it filled was a username input, or that the button it clicked submits a form. It's running commands against a stateless API. The moment the page structure changes, a new modal shows up, an element gets renamed, there's an extra redirect, the script breaks. Not because you got the logic wrong. There just wasn't any logic. Just steps.
This is where browser automation sits in 2026. Playwright, Selenium, Browser-use, Stagehand, they all give you good primitives. None of them give you a model for what a browser session actually is. State lives in local variables. Context evaporates between function calls. The relationship between a session, its pages, and their elements exists only in your head.
I built jac-browser to try to fix this. It's a browser automation library written entirely in Jac, 153 walkers that wrap Playwright. But the wrapper isn't really the point. The point is that browser sessions are graphs, and Jac gives you graphs as a first-class language construct. The result is a library where the data model is the automation framework.
Every model demo you've ever watched -- Anthropic, OpenAI, whoever -- follows the same script. Someone types a request, the model produces something impressive, the audience applauds. It's a compelling showcase for the model. It's a terrible workflow for an engineer.
I've spent years in compilers, language runtimes, and distributed systems -- the kind of work where handwaving doesn't survive contact with reality. When AI coding assistants arrived, I fell into it like everyone else. The demos make it look so clean.
It took longer than I'd like to admit to recognize the pattern: it's the same thing you see when a developer copies from Stack Overflow without understanding the code. The demo format trains passivity. It optimizes for throughput at the expense of understanding. That tradeoff doesn't hold up in production.