Make Grok Bot Use Less: 9 Free Tools to Install

Me too!

Free PDF · 8 pages

The guide as a PDF, for you and for your agent.

Read it at your pace, print it or hand it to your AI agent: the prompt to paste is already inside.

Download the PDFCopy the link for your agent

If your weekly Grok Bot usage runs out in three days, the problem is often not how much you work, but how much text the Bot reads for each job. Nine free tools do the dirty work first: they clean the page, trim the file, find the right line. The Bot reads only the result. You don't need to code: the Bot installs them once, in half an hour, and then uses them on its own.

Before You Start: 5 Words to Know

• Token: the Bot's "fuel". Every word the Bot reads or writes uses a few tokens. An ordinary web page can be worth tens of thousands of tokens, even if you only care about three lines.
• Weekly usage: the tank of tokens your plan gives you each week. When it's empty, the Bot stops until the reset, unless you move to a higher plan.
• Repository (or repo): a project's public folder on GitHub, the site where developers share their work. All nine tools in this tip are free, open source repos.
• The Bot's computer: Grok Bot doesn't work on your PC but on its own cloud computer. The tools get installed there: nothing changes on your machine.
• Skill: a saved instruction the Bot follows every time, without you rewriting it. You open it by typing / in the chat.

What It Is

Everything you make Grok Bot read is paid for in tokens. A web page is not just the article: it has menus, ads and hidden code. A spreadsheet has thousands of rows when you need one number. These tools sit in the middle: they take the big file, keep the useful part and pass only that to the Bot.

Think of an assistant who, instead of reading you a whole binder out loud, brings you just the page with the answer. Same job, far less effort.

Why It Matters

In the video this tip is based on, Jordan Rafealov ran seven real jobs twice: once the normal way, once through the tools. The normal way used almost 4.5 million tokens; with the tools, 57,900. That's the "98.7% less" in the video's title.

Read that number carefully: almost all of the saving comes from a single test, a huge spreadsheet. On the other jobs the cut ranges from about a third to 99%. Still a lot, and on things you do every week: reading pages, PDFs, lists and spreadsheets.

There's also a rule in the Grok Bot docs: every routine run spends usage, and an hourly or short-interval routine can use a week of usage in a day. If the Bot reads whole files on every run, the bill multiplies.

How to Do It

1. Open a conversation with your Bot. Your usual one is fine.
2. Paste the prompt below. The Bot installs the tools on its computer and tells you, line by line, whether each one worked. If one fails, ask it to retry or skip it: the others still work.
3. Let it save the rules as a skill. That's the part of the prompt that says "save these rules". From then on the Bot knows which tool to use for each kind of file.
4. Work as before. Ask for things the normal way. The Bot runs the tools before it reads.
5. Check after a week. See whether your weekly usage lasts the week. If a rule doesn't fit your work, ask the Bot to remove it.

Ready to copy. Paste it into a conversation with your Bot:

1. Trafilatura: Web Pages Without the Clutter

What it does: keeps only the article text of a web page and drops menus, ads and code.

In the video's test: the Costco Wikipedia page went from about 400,000 tokens to 26,000.

Use it for: news, product pages, competitor sites, any page you want summarized.

How to ask: "Summarize this page in 5 points: [link]".

2. jq: Only the Fields You Need

What it does: when a site replies with a long block of data (JSON, the format sites use to pass information around), it keeps only the fields you ask for.

In the video's test: the name, stars and language of nine GitHub projects cost 174 tokens instead of almost 18,000.

Use it for: price lists, order lists, contacts: keep prices, totals, names and emails, and leave the rest.

How to ask: "From this list, give me only name, price and availability".

3. ripgrep: The Right Lines, Not the Whole File

What it does: searches many files for a word and returns only the lines that contain it.

In the video's test: searching seven code files cost about 300 tokens instead of 30,000.

Use it for: finding a customer, a number or an entry across many text documents.

How to ask: "In the quotes folder, find where Smith Ltd appears".

4. DuckDB: Questions to a Spreadsheet

What it does: asks a big spreadsheet (CSV or Excel) a question and returns only the answer.

In the video's test: a list of 86,000 airports, with the question "which countries have the biggest airports?". From almost 4 million tokens to 89. This test makes up nearly all of the 98.7%.

Use it for: "which product sold the most", "how many customers do we have in Leeds", "which store earns the most".

How to ask: "In sales-2026.csv, what were the 5 best-selling products in October?".

5. Docling: PDFs Into Clean Text

What it does: turns a PDF into tidy text, with readable headings and tables.

In the video's test: a 9-page PDF went from 12,700 tokens to just under 8,000, about a third less.

Use it for: invoices, contracts, reports, quotes in PDF.

How to ask: "From this PDF contract, tell me the term, the price and the penalties".

6. changedetection.io: Only What Changed

What it does: watches a page and reports only the differences since last time.

In the video's test: from 35,000 to 20,000 tokens, 43% less, re-reading a page it already knew.

Use it for: competitor prices, job listings, supplier pages. It's a separate service that stays on all the time, not a skill rule: ask the Bot to set it up, or use the project's hosted version.

How to ask: "Watch [competitor]'s pricing page and tell me only when something changes".

7. RTK: Shorter Command Output

What it does: shortens what terminal commands print (the terminal is the window where you type commands to a computer) and keeps the important lines.

In the video's test: a list of 30 project updates went from 5,000 to 2,399 tokens, about half.

Use it for: when your Bot works on code, websites or servers. If you only use it for email and documents, you can ignore it.

8. Tenacity: Retry Without Starting Over

What it does: inside a Python script, when a call fails it waits and retries on its own.

In the video's test: a call made to fail twice on purpose worked on the third try, in 2.2 seconds, without asking the Bot anything.

Use it for: scripts that download reports or call sites that sometimes don't answer. The saving is not having the Bot redo the whole job.

9. Pydantic: Check Data Before Using It

What it does: checks that a result has the right shape: fields present, numbers where numbers belong.

In the video's test: on a sample reply it found two errors (a badly written number and a missing link). You fix those two and redo nothing else.

Use it for: contacts without an email, orders without a price, data going into a sheet.

Short on Time? Start With Three

For a small business or a freelancer, the jobs that burn the most usage are nearly always the same: reading web pages, spreadsheets and PDFs. To start small, have the Bot install only Trafilatura, DuckDB and Docling. Add the others when you need them.

Small Test

Pick a job you do often with a big file, such as a sales CSV or a long PDF. Before installing the tools, ask the Bot your question and watch how much your weekly usage moves. Then paste the prompt above and ask the same question on the same file. If the Bot says it used duckdb or docling and usage moves much less, it works.

Frequently Asked Questions

Do I need to code? No. You paste a prompt and the Bot does the rest. Tenacity and Pydantic only matter if the Bot writes programs for you.

Does it cost anything? The tools are free. Installing them uses a bit of usage, once.

Can I break something? The tools live on the Bot's computer, not yours. At worst one fails to install and the Bot works as before.

What if the Bot doesn't use the tools? Remind it: "use the read less skill". Or type / in the chat and pick the skill.

Does it work with other assistants? Yes, with those that can use a terminal, such as Claude Code or Codex. Not with an assistant that only lives in the browser.

Watch Out For

• The numbers are the video's, not ours: they come from the author's tests, on seven jobs he picked. In the "before" runs the Bot read the whole file. If you already work in a tidy way, you'll save less.
• The 98.7% comes almost entirely from DuckDB. On pages, PDFs and commands, expect a cut between a third and 95%.
• The install prompt is ours. In the video, the author's prompt sits in his community. The one above does the same job and is free.
• An extract can miss a detail. If an answer looks incomplete, ask the Bot to read the whole file that one time.
• Limits and plans change: check the Grok Bot docs and your account settings.

Next step: pick the right model for each job with cost-aware routing, and give the Bot a memory so it stops repeating mistakes.

Source: the video 9 GitHub Repos That Cut Grok Bot Usage by 98.7% by Jordan Rafealov. Repositories: Trafilatura, jq, ripgrep, DuckDB, Docling, changedetection.io, RTK, Tenacity, Pydantic. No affiliation.

In short: Grok Bot burns usage when it reads whole files: have the data cleaned and trimmed first, and the Bot reads only what it needs.