Pythia

Ask a question. Get a report.

Pythia is a conversational reporting app for FileMaker databases. Point it at any FileMaker Server file, pick the tables, and ask questions in plain English. It answers with real numbers, charts, and printable reports. No layouts to build, no relationships required, no plugins.

Nothing about your schema is baked in: Pythia reads the connected database's structure at runtime, and the dashboard and starter reports are generated from your data.

Install

One line. Mac or Linux, any terminal:

curl -fsSL https://navarre.ai/get/pythia | sh

Windows, plain PowerShell (no WSL, no admin):

irm https://navarre.ai/get/pythia.ps1 | iex

It installs the Fly CLI if needed, deploys Pythia to your own Fly.io account (Fly builds it in the cloud; no Node, no Docker, nothing else on your machine), and opens your browser. Everything else, the FileMaker connection, the AI key, the tables, happens in the browser.

You'll need: a hosted FileMaker database with OData enabled, a Fly.io account (a Pythia instance runs a few dollars a month), and an Anthropic API key. No FileMaker database handy? Pythia starts with a small bundled sample dataset so you can try it with nothing but the AI key.

Anything odd during install: TROUBLESHOOTING.md covers every failure real users have hit, and ends with a free-call escape hatch. Installing with an AI tool? Point it at AGENTS.md; the whole install is one command, and a good agent run needs only a handful of approvals.

The two rules

1. FileMaker is the source of truth. The numbers come from your data, not the model. The AI writes queries; a local SQL engine runs them. The model never invents a number, and a saved report re-runs with no AI at all, so it returns the same result every time.
2. You never see the plumbing. No SQL, no UUIDs, no internal keys on screen. Answers are plain business English.

How it works

Node + Express, one npm dependency, a single-file web UI. Pythia syncs your chosen tables over OData into a local DuckDB copy on the app's own volume, so questions are fast and your server isn't hammered. With a UUID primary key and a modification timestamp on a table, syncs are incremental (only changed records travel).

What reaches the AI: your table and field names, your questions, and up to 60 rows per query result so it can read what it found. Bulk data never leaves your server, and the whole app runs in your own cloud account.

Your own copy

The source is yours to change. To customize:

git clone https://github.com/Navarre-AI/pythia && cd pythia
npm install && npm start        # local, sample data
fly deploy                      # ship changes to your own app

Want a private copy on GitHub? Don't fork (forks of public repos can't be private). Use GitHub's Import repository, or push your clone to a new private repo: git remote set-url origin <your-repo-url>.

More

• EXTENDING.md: fork it and build — the architecture, the seams, and recipes for write-back, semantic search, logging, dashboards, and UI generation.
• SETUP.md: the FileMaker Server side (enable OData, one scoped account).
• TROUBLESHOOTING.md: install first aid, from real field failures.
• GOTCHAS.md: FileMaker/OData platform quirks this code handles.
• License: free to use, run, and modify, with attribution; no resale without permission. See LICENSE. © 2026 Navarre AI (Fermata Software LLC), created by Matt Navarre. navarre.ai