Skip to main content

​
Get an inspectable answer

By the end of this guide, Fella will have scanned a folder, answered a question from its contents, and shown you the computation behind the result.
1

Install the desktop app

Download the current unsigned build from GitHub Releases. Releases provide a universal macOS .dmg, Windows installers, and x86_64 Linux .AppImage and .deb artifacts.
Current builds are unsigned. Review a script before piping it to a shell, or download an installer and compare it with the release’s SHA256SUMS.
2

Connect a tool-capable model

Fella is BYOK-only. Connect a hosted provider with your own API key before asking a question.
Run /login, choose OpenAI, Vercel AI Gateway, xAI, Ollama Cloud, OpenRouter, or a custom OpenAI-compatible endpoint, then use /model to inspect the provider’s current model list.
Hosted providers receive prompts and tool results. Keys are stored in local auth.json with 0600 permissions where supported, not in the settings database, browser storage, or transcript.
3

Open a focused folder

Choose Open folder, drag a folder into the window, or enter:
or to open the file explorer:
Start with a small folder containing only relevant files. Fella scans up to eight levels deep, skips hidden paths, does not follow links, and applies root .fellaignore name or path-prefix rules. A root fella.md supplies context and is not treated as evidence data.
4

Confirm what loaded

Run /files.
Check skipped files, inferred types, units, date shapes, nulls, and ingest notes before trusting an aggregate. This catches the most common source problems earlier than a natural-language question does.
5

Ask and inspect

Ask for a bounded outcome that names the measure and time range.
Open the working fold under the response. Confirm the right source was selected, the query filters match your wording, units are correct, returned rows support the answer, and no verification warning needs attention.

​
What the shipped app can use

SQL result materialization is capped at 1,000 rows by default and a query is interrupted after 15 seconds. These are guardrails, not target dataset sizes.

​
A good first-question pattern

However, the goal of fella is to eventually support looser question patterns such as
Specific wording reduces interpretation errors. If the answer matters, inspect the query rather than relying only on a green verification label.

​
Next steps

Use task recipes

Adapt tested prompt shapes to your own files.

Read the limits

Understand verification, Python, formats, and scale before relying on a result.