An app for processing spreadsheet data with a local LLM. Try it out now at https://decentapps.net/hone .
The basic flow of using Hone is:
- Import some data from a spreadsheet.
- Ask a question to every row of the spreadsheet with a prompt template.
- The LLM-provided answers to that question are filled into a new column of the spreadsheet.
- Export the new column to your spreadsheet.
Importing and exporting can be done from standard spreadsheet files or via the clipboard. A nice way to use Hone is to have your spreadsheet app open, and just copy and paste values between Hone and the spreadsheet app.
- Import/export Excel (XLS, XLSX), CSV, and TSV files
- Copy and paste sheets with Excel, Google Sheets, and Numbers. Other spreadsheet software probably works too
- Design and test prompt templates quickly, inserting input fields from your spreadsheet
- View sheets with up to 1000 rows and 256 columns
- Prompt execution jobs can be paused and resumed
- Llama 3.2 8b model used for executing prompts, running in your browser (Thanks, WebLLM!)
- Example data and prompts included to try it out quickly
- Developed with high-security practices (limited dependencies, no service calls, no persistence of generated data)
- Fully open source
All the data you import or create within Hone stays on your device and only when you are using Hone. There are no calls to services, including the LLM. After the app loads, you could disable your Internet, and Hone would continue to work purely offline. You must explicitly export your generated data to keep it. Otherwise, it disappears when you close the browser tab for Hone.
I don't need to assure you that I won't sell your data to 3rd parties or use it to train AI models. Because I don't have your data! Only you do.
- LLM - large language model
- prompt - a question that can be asked to the LLM that will generate an answer
- prompt template - it's a prompt with insertion points used for adding input data. In Hone, the inputs come from per-row spreadsheet values.
- sheet - tabular data with columns and rows. A single Excel file can contain multiple sheets.
- Ask a question that boils down to single, concrete answer. E.g., "Is this object an animal, vegetable, or mineral?" rather than "Analyze and classify the object."
- To keep the LLM from spewing lots of text, it's good to specify constraints. E.g., "Output one of the following words 'animal', 'vegetable', or 'mineral' that best classifies the object." The verb "output" is especially useful for focusing the LLM.
- Don't use the LLM to do calcs (SUM(), AVERAGE(), etc). Your spreadsheet software will be faster and more reliable for this.
- Only add fields to your template that are useful for influencing the LLM's response. E.g., "What is the best gift for a child named {Name} who likes {Gift Preferences}?" is not as reliable or as fast as "What is the best gift for a child who likes {Gift Preferences}?" ("Name" field omitted from latter)
- Shorter prompt templates tend to take less time to execute. Execution time is influenced by the size of your prompt (prompt template with inserted values) and the response coming back.
If you find your prompt template going beyond a few sentences, consider breaking it into multiple prompts with the output of each going into a separate column. You can reuse these outputs as inputs to a prompt that synthesizes everything. An example using the Christmas Decisions example data:
Prompt 1 - "Output a score from 1 (extremely naughty) to 100 (extremely nice) for a child with '{Feedback}' feedback." -> output to a "Moral Rectitude" column. Prompt 2 - "Output a single gift name for a child with '{Gift Preferences}' preferences that has a value of ${Moral Rectitude} USD." -> output to a "Gift" column.
We run prompt 1 to give each child a "moral rectitude" score. We then use that output as an input to prompt 2.
We could have instead combined these into a single prompt, e.g. "Output a single gift name for a child with '{Gift Preferences}' that has received '{Feedback}'. Gift value should be from $1 to $100 USD based on their feedback." But the results will be far less reliable. And we'll have less transparency on how gift decision was made.
If you are wondering if I actually want to use LLMs to evaluate children's moral rectitude, well... read the next section.
LLMs aren't magic. They're just good at predicting what text should follow some other text. The LLM used by Hone, like all the others, receives some input and generates output. It's what you, the human, does with that output that can be dangerous or harmful.
So for example, suppose you had a collection of job applications in a spreadsheet, and you asked Hone something like "Given all these inputs from the job application, output a score from 1 (never hire) to 10 (absolutely hire)." Decisions made like that could be terrible mistakes that hurt your company. You might also harm the future of deserving applicants that didn't fit the average profile of a successful jobseeker. In fact, you might even reinforce cultural biases that had become encoded into that average profile. The LLM was trained on the vast corpus of things people say on the Internet, and no matter how hard we scrub the LLM, it's just gonna reflect a lot of the mean and mistaken things people believe.
We hear a lot of talk about SkyNet-style AI takeovers and the universe being turned into paper clips. But I'm telling you, it's mainly the humans we have to worry about. And that includes you and me when we get up to something stupid. Please, pay attention to how you use LLM-generated data:
- Use prompts that ask clear questions with concrete answers. You should be able to judge for yourself if an answer from an LLM is correct or not.
- Be aware of cultural data that can exist in inputs like people's names or freeform writing and exclude from prompts unless needed.
- Put appropriate manual review on data that is used to make consequential decisions, e.g., if you're going to use LLM-generated data to judge a science fair entry, it's pretty important to confirm the data with your human eyes.
- If you're making decisions with generated data, design prompt templates to provide decision support information rather than making the decision itself.
- Avoid automating actions in the real world based on AI-generated data. Yes, you could use Hone to send a personalized newsletter to 10,000 people based on their LinkedIn profile data. But there is potential for a random hallucination to generate exactly the wrong words. And that might mean serious consequences to you - lawsuits, reputational damage, or guilt over pain you've caused.
You can use Github Issues to file an issue.
But also, I want my baby app to do well in the World. Right now, in 2025, I'm motivated to work directly with you as a brand new Hone user. Consider reaching out to me as described in the "Contacting" section below.
My code and other files in this repository are licensed under the MIT open source license.
The fonts used by the template are hosted from decentapps.net rather than included in this repo. Their licensing is separate, but will be something compatible with open source, e.g., OFS SFIL. If you want to self-host the fonts rather load them from decentapps.net, the easiest thing might be to just find the same fonts from other sources and verify the licensing for your use.
If you want to check the licensing I used for a hosted font, you can replace the filename of the URL that loads a WOFF or WOFF2 file with "LICENSE". So for example, the font served from "https://decentapps.net/fonts/hobby-of-night/hobby-of-night-webfont.woff2" was licensed to me according to terms found at "https://decentapps.net/fonts/hobby-of-night/LICENSE".
I am so excited about local-LLM web apps, that I made a project code generator for them. Basically, you can run "npx create-decent-app YourProjectName" from a terminal window, and it will give you functioning local-LLM web app that is ready to be modified to fit your vision. See my create-decent-app Github repo for more info.
The project isn't open to contributions at this point. But that could change. Contact me if you'd like to collaborate.
You can reach me via my LinkedIn profile. I'll accept connections if you will just mention "hone" or some other shared interest in your connection request.
-Erik Hermansen