For Gemini CLI

A large context window is not the same as the right context. Aestheria, the design context layer, is the part Gemini CLI cannot infer from your repository however much of it fits.

Volume is not the missing piece.

Gemini CLI can take in an enormous amount of a codebase, which makes the design problem easy to misdiagnose. Fitting the whole repository in context does not tell an agent which of four greys is the border grey, or that this screen owes the user an empty state. That information is not in the repository at any size.

What an agent reads from a large codebase is precedent, and precedent includes every shortcut anyone took. Generating from the average of your own inconsistencies is a reliable way to extend them.

Query the decisions instead of inferring them.

One command connects the library, and the agent asks for what it needs rather than being handed everything at once.

gemini mcp add --transport http aestheria https://mcp.aestheria.design/mcp

The same server can be declared in ~/.gemini/settings.json with httpUrl if you would rather keep it in config. Add --trust to skip per-tool confirmation prompts once you know what it calls.

What the agent gets back.

Tokens with the rule for each, so a colour is chosen rather than sampled. Components with their variants and states, so the existing one is reused. Page templates, so a new screen is composed the way your screens are composed. Six skills carrying the judgment between them. And validate_snippet, which answers the only question that matters at review time: does this actually follow the system.

• 405 tokens
• 6 themes
• 37 components
• 27 charts
• 63 templates

Questions

How do I connect Aestheria to Gemini CLI?

Run `gemini mcp add --transport http aestheria https://mcp.aestheria.design/mcp`, or add the server to `~/.gemini/settings.json` with `httpUrl`. The hosted server is free with an account.

My context window is huge. Why do I need this?

Because the missing information is not in the codebase. Usage rules, component states, and the judgment that picks between two valid options were never written down in a form any agent can read, whatever the window size.

Can I skip the confirmation prompts?

Add `--trust` when you register the server. The tools are read-only queries against the library plus a validation call, so there is nothing destructive in the set.

Does it work offline?

Take the kit instead of the server. The files live in your repository, so the tokens, components, and skills are read from disk with no network in the loop.