llms-furl

Furl llms-full.txt (or llms.txt link lists) into a structured tree of small, searchable files you can assemble into LLM context with standard Unix tools.

                                          ├── api/
                                          │   ├── auth.md
┌──────────────────┐                      │   └── rate-limits.md
│  llms-full.txt   │  ─ npx llms-furl ─▶  ├── concepts/
│  (400KB blob)    │                      │   ├── context.md
└──────────────────┘                      │   └── rag.md
                                          └── examples/
                                              └── sdk.md

furl /fɜːrl/ — to roll up; to make compact. A play on "full."

No vectors, no tools. Just files and bash.

Why filesystem-based context?

"The primary lesson from the actually successful agents so far is the return to Unix fundamentals: file systems, shells, processes & CLIs. Don't fight the models, embrace the abstractions they're tuned for. Bash is all you need." — @rauch

LLM agents perform well with Unix-style workflows like find, grep, jq, and pipes. Rather than stuffing everything into the prompt, you can keep large context local in a filesystem and let agents retrieve smaller slices on demand — this is filesystem-based context retrieval.

llms-furl turns context selection into a Unix problem, not a prompt-engineering problem.

Install

Requirements: Node.js >= 20.

npm install -g llms-furl
# or one-off
npx llms-furl --help

Quickstart

llms-furl https://vercel.com/docs/llms-full.txt
tree -L 3 llms-furl/vercel.com

llms-furl/
└── vercel.com
    ├── index.json
    ├── api/
    │   ├── auth.md
    │   ├── files.md
    │   └── rate-limits.md
    ├── concepts/
    │   ├── context.md
    │   ├── rag.md
    │   └── tasks.md
    └── examples/
        ├── file-upload.md
        └── sdk.md

Each file is a leaf — a small, self-contained piece of the original document, split along its natural section boundaries.

Now you can use standard Unix tools to build exactly the context you need.

# Find anything related to rate limits
rg "rate" llms-furl/vercel.com/docs

# Collect all API-related docs
fd . llms-furl/vercel.com/api | xargs cat

# Build a context for "file upload"
rg -l "file upload" llms-furl/vercel.com | xargs cat > context.txt

Pipe that directly into your LLM:

cat context.txt | llm "Summarize how file uploads work in this API"

Usage

llms-furl <input> [output-dir]
llms-furl split <input> [output-dir]
llms-furl list [output-dir]
llms-furl remove <files...>
llms-furl rm <files...>
llms-furl clean [output-dir]

Options:

• --debug, -d show split diagnostics and flattening info
• --help, -h show help
• --version, -v show version

Output layout

• URL input defaults to llms-furl/<host> (for example, llms-furl/vercel.com).
• File input defaults to the current directory unless output-dir is given.
• Output file paths are derived from each page URL (strip leading/trailing slashes and .md/.html).
• If all pages share one top-level directory (for example, docs/), llms-furl flattens it for URL inputs (shown in --debug).
• index.json is written alongside the output and contains a tree plus source (and name for URL inputs).

Split patterns

llms-furl detects common llms-full formats automatically:

• Pattern A: # Title followed by Source: https://...
• Pattern B: <page>...</page> with frontmatter containing source_url
• Pattern C: # Title, blank line, then URL: https://...
• Pattern D: Vercel-style dash-separated blocks with title: and source:

Code blocks are ignored when detecting boundaries.

llms.txt link lists

If a site only provides llms.txt, llms-furl reads every link in the list, fetches each page, and writes each one as a leaf file. Relative links are resolved against the llms.txt URL.

llms-furl https://cursor.com/llms.txt

Integration hints

When output is inside llms-furl/, llms-furl maintains llms-furl/AGENTS.md and may offer to update:

• .gitignore to ignore llms-furl/
• tsconfig.json to exclude llms-furl
• AGENTS.md to add a llms-furl section

In TTY, you get a y/n prompt; in non-interactive runs it prints hints only. Consent is stored in llms-furl/.llms-furl.json.

llms-furl lets you treat your LLM documentation the way Unix always wanted you to: as a living, searchable filesystem of knowledge.

Acknowledgments

This project was inspired by opensrc. Thank you for the great idea!