fella

The more an AI can do for you, the more it can do to you.

General-purpose agents are built to take actions. Fella is built to help you understand: ask a question, investigate the relevant context, and get an answer whose basis you can inspect.

Your computer is full of data that never became insight: a year of expenses, a workout log, project exports, PDFs, notes, and files with names you no longer remember. Fella lets you ask about those files without first building a data warehouse or learning SQL. A workspace organizes local evidence; it is context for the conversation, not the product itself.

Fella combines a Rust analytics engine with an opinionated AI harness. The model drives interpretation, decomposition, tool choice, and explanation inside a reviewed, read-only boundary. The local engine catalogs files, queries data, searches documents, makes charts, compiles supported analytical plans, and checks computational claims. General questions can be asked without mounting a folder. Visible web research when external sources matter remains future work; the current release does not send questions or local data to a web-search service.

For analytics, a plausible answer is not good enough. The goal is to reach the right answer the first time, with as little unnecessary reasoning between the question and result as possible. The model does not need file writes, shell commands, desktop control, or the ability to act on your behalf. When a question benefits from outside information, the intended route is bounded, read-only, and visible—not an unrestricted browser agent.

The result is an opinionated system: a model-led analysis loop with reviewed, read-only tools, deterministic execution where exact computation matters, and clear limits when the available evidence cannot support a claim.

Three things Fella optimizes for

Fella does one useful thing deeply: it makes the personal data already on your computer understandable without making the reasoning needlessly complicated.

Consistency	A coherent model-led loop and reviewed tools keep the route from question to answer understandable.
Correctness	The engine computes the numbers, the harness checks the work, and the result should be right the first time whenever the data can support it.
Efficiency	Deliberate context and proportionate tool use keep wasted reasoning down without an arbitrary tool-count limit.
Read-only analysis	Model-facing analysis tools can inspect and compute over mounted files, but cannot alter source files, run shell commands, or take external actions.
Useful on real files	Mount the folder when a question depends on it; ask in plain language and get an analysis or chart without building a warehouse or learning SQL first.
Bring your own model	Connect OpenAI, Vercel AI Gateway, xAI, Ollama Cloud, OpenRouter, or a custom OpenAI-compatible endpoint.

From a question to an answer

your folder, as it is
        ↓
bounded workspace + typed local tables
        ↓
model → observations · SQL · document search · bounded Python · charts
        ↓
patterns · comparisons · charts · caveats
        ↓
a clear answer you can use

The model is the analytical driver, not the source of truth for the numbers. It can choose and revise a route through the data, but it reaches the workspace only through bounded read-only tools. Fella's Rust analytics engine performs the computation, optionally compiles a semantic plan, and a separate verification pass checks the result against real data. If the folder cannot support the claim, Fella should say so.

When you want to go deeper, the result opens up: source files, calculations, and verification are there to inspect. Evidence supports the analysis without getting in the way of using it.

That makes Fella useful for questions such as:

• “Where did my spending change over the last six months?”
• “How does my sleep relate to my workout days?”
• “Which projects have the most unresolved work?”
• “Show me the trend, and let me see the records behind it.”

Supported files

Type	Formats	Behavior
Delimited tables	.csv, .tsv, .tab	Detects column types and loads a queryable local table.
JSON tables	.json, .ndjson, .jsonl	Reads JSON objects, arrays of objects, or newline-delimited records as tables.
Spreadsheets	.xlsx, .xlsm, .xlsb, .xls	Loads each readable worksheet as a table.
Documents	.pdf, .txt, .text, .md, .markdown, .log	Searches extracted text directly; PDFs need a selectable text layer. No OCR or embedding index is required.
Parquet	.parquet, .pq	Available only in a custom DuckDB build; not included in release builds.

Install

Download the latest installer from the releases page, or use the platform installer:

macOS / Linux

curl -fsSL https://lilfella.app/install.sh | sh

Windows PowerShell

irm https://lilfella.app/install.ps1 | iex

Upgrading from the published 0.2.0 Tauri app: install the Electron build manually from the releases page once; /update in the old shell cannot convert the app. The new shell reuses the existing dev.fella.app data directory, so settings, provider credentials, and conversation history remain available. After that migration, /update checks and applies Electron releases.

To get started:

1. Type /login to connect a model provider with your own API key, unless one is already configured.
2. Choose a model with /model.
3. Ask a general question right away, or mount a folder with /open when your question depends on local files.

Fella does not require a Fella account. Provider credentials are stored locally in auth.json, separate from settings and conversation history.

Using the app

The main surfaces are intentionally small:

• Ask — start a conversation or open another tab.
• Search — search across conversations, files, repositories, and app resources.
• Repositories — folder mounts with their conversations and workspace view.
• Projects — optional, user-created knowledge pages associated with a repository.
• Workspace — inspect sources or edit the user-authored fella.md context file.
• Settings — provider, model, appearance, and analysis preferences.

Useful commands:

Command	Purpose
/open	Mount or switch to a folder.
/files	List the files and tables Fella found.
/schema <name>	Inspect a table's columns, types, and null rates.
/sql <query>	Run a read-only SQL query directly.
/context	Open the workspace's fella.md guide.
/memory	View or clear learned notes for the current folder.
/history	Browse and reopen saved conversations.
/update	Explicitly check for and install a verified update.

Ctrl becomes Cmd on macOS. The command palette is available with Ctrl/Cmd+K.

Privacy and safety

• Analysis tools read the mounted workspace but cannot modify its files or run shell commands. You can edit the user-authored fella.md context file in the app. Settings, conversations, and memory are stored in Fella's local app data.
• The current release has no shell, desktop-control, or web-research tools. Web research remains planned as a bounded, visible, read-only route; it must not send mounted-file contents, rows, snippets, workspace memory, or credentials to a search service.
• Provider API keys stay in local auth.json with restrictive permissions where supported; they are not stored in the settings database or browser storage.
• With a hosted provider, the question and relevant tool results are sent to that provider. Built-in providers are hosted; a custom OpenAI-compatible endpoint can be configured to point to a local service.
• Updates happen only when explicitly requested with /update.

Read the security policy and product commitments.

Build from source

The desktop shell is Electron, with the SvelteKit UI and Rust analytics engine running as a local sidecar. See docs/DEVELOPMENT.md for toolchain requirements and provider setup.

git clone https://github.com/Avijit-Kumar-GIT/fella.git
cd fella
pnpm install
pnpm electron:dev

Useful verification commands:

pnpm check
pnpm build
cargo test --manifest-path backend/Cargo.toml --locked

The Electron shell and its line-delimited JSON bridge are described in docs/ARCHITECTURE.md.

Documentation

• Documentation guide — product, architecture, design, contributor guides, and dated audit records
• User documentation — Mintlify content
• Contributing — contribution guidelines

Status

Fella is an early, pre-1.0 release. The core personal analytics workflow is implemented; interfaces and supported providers will continue to evolve.

License

MIT