Your Nonprofit's Data Agent — starter template

A minimum viable data agent. You ask a question in plain English about your spreadsheets; the agent writes a small query, runs it, and answers in plain English.

This is the companion code for the guide at ourcommunity.tech/build-your-own-data-agent.

Built by Our Community Tech — a nonprofit that helps other nonprofits with technology.

---

What you'll need

• A free Replit account (no local install needed)
• An OpenAI API key (free to create; pay-as-you-go after that — typically $2–$5/month at low volume)
• About 30 minutes

That's it. You do not need to know Python.

---

Quick start in Replit

1. Fork this template (click "Use template" on the Replit page).
2. Add your API key as a secret. In the left sidebar click Secrets, add a secret named OPENAI_API_KEY, and paste your key.
3. Click the big green "Run" button. The sample data loads and you get an interactive prompt.
4. Ask a question. Try one from the list below.

---

Sample questions that work well on the included data

• Which donors gave less this year than last year, and by how much?
• Who lapsed this year (gave last year, nothing this year)?
• What was our average grant size by program?
• Which programs are over budget, and by how much?
• Which funders gave us more than one grant?
• What's our total grant revenue by status?
• Which cities do our top 10 lifetime donors live in?

---

Using your own data

1. Export your data to CSV. Each table should be one CSV file. Spreadsheets → File → Download → CSV works fine.

2. Drop your CSVs into the /data folder in this Replit.

3. Open agent.py and edit the DATA_FILES dict near the top:

   DATA_FILES = {
       "donors": "data/my_donors.csv",
       "programs": "data/my_programs.csv",
   }

   The keys are the names the agent will use internally (keep them short and clean — donors, grants, budget). The values are the paths.

4. Click Run.

Important: before you point this at your real data, read SAFETY.md. There are kinds of data this tool is not appropriate for.

---

What's in here

File	What it does
agent.py	The whole agent — about 250 lines, heavily commented. Start here.
generate_sample_data.py	Recreates the sample CSVs. You do not need to run this.
data/donors.csv	60 synthetic donors with giving history.
data/grants.csv	24 synthetic grants across programs and funders.
data/program_budget.csv	7 synthetic program budgets with YTD actuals.
SAFETY.md	What you should and should not put through this. Please read.
requirements.txt	Python packages (Replit installs these automatically).

---

How it works (one paragraph)

The agent loads your CSVs into pandas, sends the model a summary of your columns (not your actual data), asks it for a small pandas snippet that answers your question, runs that snippet in a narrow sandbox, and sends the result back for a plain-English write-up. Two model calls per question. At gpt-4o-mini prices, most questions cost a fraction of a cent.

---

What it can't do (honestly)

• It is not real-time. It only knows what's in your CSVs at load time.
• It is not for huge data. This minimal version comfortably handles a few hundred thousand rows. Beyond that, move to a real BI tool.
• It is not a replacement for a data analyst. It helps when you don't have one.
• It will occasionally be wrong. Check the numbers on important questions. Toggle /code at the prompt to see the query it ran.

---

Where to go next

Once you have this running on your own data, the obvious next steps are:

1. Connect a donor database (Salesforce, Bloomerang, Little Green Light) so you do not need to export CSVs each month.
2. Combine multiple sources (grants + accounting + program outcomes) in one agent session.
3. Schedule it. Run a fixed set of questions every Monday morning and email the results to your team.

We are writing follow-up guides on each of those. If you want to be the first to see them, or if you need hands-on help, drop us a note at ourcommunity.tech.

---

License: MIT. Use it, fork it, remix it, ship it for your mission.