AI tools are better when they can look something up instead of guessing. This project downloads official documentation onto your computer, keeps those copies up to date, and lets you (or an AI assistant) search them without going online every time.
Think of it as a filing cabinet. You choose which websites to save. The
copies live in a folder called reference/ on your machine. This
repository is only the toolkit that fills that cabinet — the downloaded
pages are created on your computer, and they are not stored in git.
You do not need to be a programmer. You can pick numbers from a menu, or hand this file to an AI assistant and let it do the work.
Open this project (or this README) in Claude, Codex, Cursor, Copilot, or any similar assistant, and paste:
Set up this documentation toolkit on my computer and download the starter docs.
Follow the section "If you are an AI agent" in README.md.
Do not copy files to other machines unless I ask.
If you also want a particular website saved, add its address to that message. Then let the assistant run the commands. You should not need the menu.
You are setting this toolkit up for a human. You are not developing it.
Do not edit source, tests, or gitignored config. Do not run the
interactive menu (./bin/scanner) — it waits for keyboard input.
-
Confirm
python3 --versionworks. If it does not, tell the human to install Python 3 and stop. Do notpip installanything. This project uses the standard library only. -
If this folder is not already
ai-reference-docs, clonehttps://github.com/theglove44/ai-reference-docs.gitand work inside it. -
Default job — fill the starter library:
python3 scripts/update_local_docs.py --only claude-code --only openai-codex --only pi
Download only those three. Do not run a bare
python3 scripts/update_local_docs.py— this machine may have extra private sources, and they are not part of first-time setup. Do not copyconfig/sources.local.jsonfrom anyone else.That hits live documentation sites and can take several minutes. A failed source leaves its previous copy in place.
-
If the human named one starter source, refresh only that one:
python3 scripts/update_local_docs.py --only SOURCE_ID
-
If the human gave a documentation URL, pre-check it, then run the command it prints (
Next:):python3 scripts/precheck_url.py --json THE_URL
Use
--sourcethey chose, or the suggested name. Keep--max-pagesat the pre-check default. Do not crawl a whole domain. If the report saysrendered_browserand there is no Chrome, Chromium, or Edge, stop and say so. If there is noNext:command, stop and explain what the report recommended — do not invent a scrape, and do not overwriteconfig/sources.local.json. -
Success is folders appearing under
reference/. List them. Tell the human they can search withrg "topic" reference/<name>/and that those files must not be edited by hand.
Never run scripts/push_reference_docs.py or
scripts/refresh_and_push.py unless the human clearly asked to copy
docs onto other machines. Never commit config/deploy_fleet.json,
config/sources.local.json, or anything under reference/.
AGENTS.md is for people changing this project. Ignore it during setup.
- A Mac or Linux computer
- An internet connection
- Git (to copy this project onto your computer)
- Python 3 (already present on many Macs)
You do not install extra Python packages.
Chrome, Chromium, or Edge is only needed later, for a few modern sites that do not show their text without a browser.
Copying the filing cabinet to other computers is optional. Skip that until this computer is working.
Check Python in Terminal:
python3 --versionIf that fails, install Python 3 from python.org.
On a Mac, open Terminal (press Command-Space, type Terminal, press Return).
- Copy this project onto your computer:
git clone https://github.com/theglove44/ai-reference-docs.git
cd ai-reference-docsIf you already have the folder, skip the clone and cd into it instead.
- Start the menu:
./bin/scannerIf that does not run:
python3 scripts/interactive_menu.pyYou should see a numbered list. Type a number and press Return. That is the whole interface.
You are set up when the menu appears. The reference/ folder is still
empty until you download something.
Right after a clone, nothing has been saved yet. In the menu, choose Update existing docs, then Refresh all managed sources.
That downloads the starter set that ships with the project: Claude Code, OpenAI Codex, and Pi. Anything else is optional — you add it yourself. It can take a while. If one site fails, the last good copy of that site is left alone.
When it finishes, open the reference/ folder. You should see one folder
per product, full of readable files.
Back at the menu, choose Scrape a new URL (“scrape” here just means download pages and save them as files).
- Paste the address of a documentation page.
- Answer the questions. The menu checks the site and picks a method.
- Give the copy a short name, like
my-product. That becomesreference/my-product/.
A longer walkthrough is in docs/scrape-your-own-docs.md.
Search the copies on your computer. If you have ripgrep
installed (rg), this is the usual way:
rg "authentication" reference/claude-code/Finder works too: open reference/ and search inside it.
Point your AI tools at the reference/ folder and tell them to search
there first. Do not edit those files by hand. Download them again when
you want a fresh copy, so the next update can rebuild the same result.
Only do this if several machines should share the same local copies, and those machines are already on your Tailscale network.
- In the scanner menu, choose Onboard a new machine.
- Pick the computer from the Tailscale list, confirm the login user,
and let it copy
reference/across.
That writes config/deploy_fleet.json on this computer (keep it out
of git) and copies the files in reference/. The other computers do not
need your private source list (config/sources.local.json).
You can still create the fleet file by hand from
config/deploy_fleet.example.json if you prefer. Later copies use
Push local docs to other machines.
Details: docs/device-deployment.md.
The scanner will not start. Make sure Terminal is inside the
ai-reference-docs folder, then try
python3 scripts/interactive_menu.py.
Python is missing. python3 --version should print a number. If it
does not, install Python 3 and open a new Terminal window.
A download fails. Try that source again later. The previous copy stays on disk.
A site comes back almost empty. Install Chrome (or Chromium or Edge) and try that site again. A few documentation sites need a real browser.
You would rather type commands than use the menu. Use the short command list in docs/quick-use-guide.md. The menu is still the easier first stop.