Skip to main content

Know what Fella does not promise

Fella is an early, pre-1.0 personal analytics app. Its narrow scope makes the normal path inspectable, but it does not remove model, data, or operating-system risk.

Correctness and evidence

For consequential decisions, validate the source and query independently. Fella is not a substitute for professional financial, medical, legal, or safety advice.

Data and formats

  • Shipped builds support delimited/JSON tables, Excel workbooks, and text-layer documents listed in Getting started.
  • There is no OCR. Image-only PDFs and standalone images are not readable data.
  • Word, OpenDocument, email, HTML, XML, archive, media, and database files are not supported inputs.
  • JSON must be table-shaped enough to ingest; arbitrary deeply nested documents may not become useful columns.
  • Excel formulas are not recalculated by Fella; workbook parsing depends on stored values the parser can read.
  • Parquet requires the optional DuckDB backend. DuckDB is not part of published installers or CI and has known ingestion-parity gaps relative to SQLite.

Scale and output

Fella targets MB-scale personal folders and everyday lookups, aggregates, comparisons, and small joins. It is not a warehouse, distributed query engine, notebook, ETL system, or multi-user server.

Security boundaries

Base file and SQL tools

The model has no built-in write, move, copy, or delete tool. Document reads resolve through the workspace catalog, and SQLite queries use a read-only connection plus a single-statement guard.

Python

Python runs without a per-call permission prompt inside an embedded wasm32-unknown-unknown RustPython guest under Wasmi. It has no filesystem, network, environment, or subprocess capability. The host provides captured output and bounded read-only SQL, and applies fuel, memory, stack, source, output, row, and response limits. The guest contains the core language and Fella’s small analytics helpers; it does not contain pandas, NumPy, SciPy, or a package installer.

Context

The user may edit fella.md in the Workspace surface. This is a human-driven write path, not an agent tool. The agent can read the file as context, but it cannot create or modify it.

Experimental MCP

/mcp is inert in the personal release. The default build has no MCP client, connector authentication, dynamic tools, or outbound connector path. A fork that adds a connector must review its permissions and network behavior separately.

Privacy and network activity

  • Fella is BYOK-only: the provider you connect receives prompts and tool results for each question.
  • /update contacts GitHub and downloads an installer only when the user invokes it.
  • /mcp introduces no network activity. The embedded Python guest has no network capability.
  • Fella does not proxy hosted model requests through a Fella service, but cannot control a provider’s logging, retention, or training policy.

Releases and platforms

Published builds are currently unsigned, so macOS and Windows can show unknown-developer warnings. Integrity depends on HTTPS and SHA256SUMS, not code signing or notarization. macOS requires 10.15+, Windows requires 10+, and Linux requires WebKitGTK 4.1. Release automation builds macOS universal plus Windows and Linux artifacts on GitHub-hosted runners, but every apply/update path has not been manually exercised on every OS.

Experimental compatibility boundary

The pack and connector designs are archived in the repository for future forks. They are not part of the default runtime. See Experimental MCP for the exact inert command and reopening criteria.

Benchmarks

Published scores are dated measurements on fixed synthetic batteries, providers, models, prompts, and grading code. They do not predict correctness for your folder. The 2026-09-10 64-case comparison and 2026-09-12 16-case hard run should be rerun after material harness, fixture, model, or endpoint changes.

Next steps

Work within the limits

Use bounded recipes and inspect the resulting queries.

See evaluation methods

Learn what deterministic tests and live model evaluations each establish.