Anonyfox/lfm

Sell before you build: one binary and a skill that make your coding agent research real buyers and
ship one checked, fast landing page with a payment link.
sh
curl -fsSL https://lfm.cash/install.sh | sh
lfm is an agent-first page compiler. Your coding agent researches real buyers, writes a design
concept from what they say, and builds one HTML template with its config and assets. lfm compiles that into a
standalone static folder and refuses to ship a page that is broken, slow, inaccessible, invisible to
search or full of generic AI design defaults. Every finding says how to fix it.
No framework, no platform, no runtime: one binary, one page, one folder you upload anywhere. No MCP
server either: a CLI with --json plus one SKILL.md is the whole agent interface.
## Ask your agent
Give any coding agent (Claude Code, Codex, Cursor, OpenCode, Gemini CLI and others) this request:
> Make a landing page for my sourdough course with lfm: https://lfm.cash/SKILL.md
SKILL.md tells the agent how to install lfm, set up a project and follow the process.
To have every agent session pick it up by itself for requests like "generate me a landing page", install
it once as a user skill:
sh
for d in ~/.agents/skills/lfm ~/.claude/skills/lfm; do mkdir -p "$d" && curl -fsSL https://lfm.cash/SKILL.md -o "$d/SKILL.md"; done
~/.agents/skills is read by Codex, Cursor, OpenCode, Gemini CLI and GitHub Copilot; ~/.claude/skills
by Claude Code. Keep only the one your agents need.
On Grok Bot, add the lfm.cash bot: it installs lfm on its own computer and runs the same
process, asking you only what you decide. Its setup is in docs/grok-bot.md.
## Install
The script at the top downloads the release for macOS or Linux (amd64
or arm64), checks it against the release checksums and puts lfm in ~/.local/bin, without sudo.
LFMVERSION=v0.0.2 pins a release and LFMINSTALLDIR changes the folder. With Homebrew:
sh
brew install anonyfox/tap/lfm
With Go 1.26 or newer:
sh
go install github.com/Anonyfox/lfm/cmd/lfm@latest
Optional: Chrome or Chromium for lfm compile --release, and Lighthouse for its scores. Missing tools
are reported and skipped, never installed.
## Start a page
sh
lfm init sourdough-course # a starter project that compiles cleanly
cd sourdough-course
lfm skills install # the process skills, where your agent loads them
lfm compile # findings, output/ and the next step
Then ask your agent for the page. It works in three phases, each with its own skill:
1. research: real buyer voices and competitor pages, verbatim and dated, into research/.
2. design: concept.md from the research: buyer, mechanism, direction, palette, type, imagery,
motion and copy, every decision cited to a source.
3. implementation: the page from concept.md, until lfm compile --release is clean and every
saved frame holds up on phone and desktop.
The compiler checks each handover: missing research, an uncited concept or a page that ignores the
concept's palette, type and imagery show up as findings. Every report ends with the next step, such as
Next (design): write concept.md from the research., and names the skill for it, so an agent working
unattended stays on the path. When it says deploy, upload output/ to the URL in lfm.json.