A coding assistant such as Claude Code, Codex, Cursor or GitHub Copilot writes better Suprnova code when it reads the Suprnova manual instead of guessing from Laravel or from older Rust web frameworks. This chapter shows you how to give an assistant the manual as plain text, how to install the Suprnova language server, and how to point an assistant at your project.
Read the manual as Markdown
suprnova.app serves every manual page twice: as
HTML for you, and as its Markdown source for an assistant. Add .md to
a page's address to get the source, with no page layout and no scripts:
The HTML page names its Markdown form in a
<link rel="alternate" type="text/markdown"> tag, so an assistant that
fetches the HTML can find it.
Two more files follow the llms.txt convention:
https://suprnova.app/llms.txtis an index: a short summary of Suprnova, then a link to the Markdown form of every chapter, grouped as the manual's table of contents groups them.https://suprnova.app/llms-full.txtis the whole manual in one document, every chapter inlined in table-of-contents order. It is large (about 2 MB), so give it to an assistant that can hold it, or let the assistant fetch the chapters it needs fromllms.txt.
Tell your assistant where they are. For example, in a conversation:
Read https://suprnova.app/llms.txt, then fetch the chapters you need as
Markdown before you change routing or models in this project.
The manual is also in the framework's repository as plain Markdown, under
manual/. An assistant that has the repository checked out can read the
files directly.
Install the language server
The Suprnova LSP is a language
server for Suprnova applications. It reads the API that #[model]
generates from rustdoc JSON, so hover, completion and type inference work
on generated methods such as Post::query(), and it finds your
application's models on its own. It runs without rust-analyzer. An
assistant that drives your editor, such as Copilot or Cursor, sees the
same completions you do.
To install it in VS Code:
- Open the releases page
and download the
.vsixfile for your platform, such assuprnova-lsp-0.3.0-linux-x64.vsix. Each one bundles its server. - In VS Code, run the Extensions: Install from VSIX... command and
choose the file, or run
code --install-extension suprnova-lsp-0.3.0-linux-x64.vsix. - Reload VS Code.
The extension ID is eas4ai.suprnova-lsp, and its settings start with
suprnova-lsp.. It exports the model API again when your code changes,
which needs the nightly Rust toolchain its settings name. The releases
are pre-releases. Each release also carries the server alone, as a
.tar.gz file for each platform, for editors other than VS Code.
Point an assistant at your project
An assistant works best when it knows how your project is built and how
you check a change. Most assistants read an instruction file at the root
of the project: AGENTS.md for Codex and several others, CLAUDE.md for
Claude Code. Suprnova does not generate one; write it yourself and keep it
short. For an application made with suprnova new, it can say:
- ----
Name the commands your project uses. The CLI reference lists
every suprnova command, and Testing shows how the tests of
an application run.
Why Suprnova diverges
Laravel's AI chapter is built around Laravel Boost, an MCP server
installed with Composer that gives an assistant tools to inspect a running
Laravel application, a documentation search API, and generated guideline
files. Boost is Laravel tooling, for PHP applications, and Suprnova does
not ship a port of it. In its place, Suprnova offers the manual as
Markdown at /docs/<chapter>.md, the llms.txt and llms-full.txt
files, and the Suprnova LSP for editor-level knowledge of your models.
The instruction file that Boost generates is one you write yourself.
Next
- Development - the local workflow an assistant runs commands in
- CLI reference - every
suprnovacommand, for your instruction file - Testing - how to check a change before you accept it
