Skip to main content
Fella reads and computes on your computer, but it is not a local-model-only application. When you use a hosted model, selected parts of the workspace may leave the computer as part of the request. The provider and model you choose determine that service’s handling of the data.

Local file access

You choose the workspace folder. Fella scans and reads supported files there; supported tabular rows are normalized into a local SQLite database in Fella’s application-data directory. Text and PDF documents are read locally by the document tools. The agent has no direct filesystem API, and the built-in tools cannot write, move, or delete workspace files. SQL accepts one SELECT or WITH statement and runs through a read-only connection. The user may explicitly edit fella.md from the Workspace context view; the model cannot write it. Analysis-derived knowledge is stored locally with its provenance and scope, and proposed facts are not silently applied as workspace-wide rules.

What is sent to a model provider

When configured, the selected provider receives the API key for authentication, along with the model requests needed to answer. Requests can include the question, recent conversation context, system prompt, registered tool schemas, workspace schema/sample rows, user or folder context, and results returned by tools. Those tool results may contain document passages (or the complete text of a short document), selected table rows, aggregate values, or Python output. Fella does not upload the workspace folder as a bulk file, but it can send file-derived text and data to the provider when the model requests and uses them. A question may cause multiple provider requests because the model can inspect, calculate, and revise in separate turns. The current runtime has no web-search or page-reading tool. It does not send workspace content to a search service, and it must not imply that it researched current web sources. /mcp is inert. Fella does not proxy model requests through a Fella service and does not transmit the local run log as telemetry.

Credentials

Provider API keys are stored in auth.json in Fella’s application-data directory, not in the settings database, browser localStorage, or conversation archives. On Unix-like systems Fella sets file mode 0600. On Windows the file inherits the permissions of the per-user application-data directory; Fella does not currently use Windows Credential Manager or the macOS Keychain. Secret-bearing commands are redacted from the transcript.

Local retention and deletion

Fella has no Fella-hosted account or server-side conversation store. Local data remains in the application-data directory until you delete it or remove that directory; there is no automatic age-based cleanup. That directory can include:
  • auth.json: provider keys.
  • fella.db: settings and local app state.
  • analysis.db: the currently mounted workspace’s normalized tabular data.
  • conversations/: saved transcripts and embedded evidence.
  • analysis/turns/: canonical analysis records, including questions, results, and detailed evidence.
  • memory/ and signals.jsonl: local folder memory and content-minimized runtime friction signals.
Deleting a conversation removes its transcript and linked Investigation. Saved analytical outputs, reports, their revisions, and the analysis records needed to support saved work remain in the workspace. The confirmation dialog describes this retention before deletion. Deleting a workspace entry removes its conversations but does not delete the folder or source files; reconnect the folder to reopen retained reports and results. Forget memory removes learned folder memory; it does not delete the workspace, provider settings, or API key. To clear all Fella-local state, quit the app and remove its dev.fella.app application-data directory. This also removes saved credentials, settings, conversations, reports, analysis records, and the local workspace database; it does not modify source files in your workspace. Default locations are %APPDATA%\\dev.fella.app on Windows, ~/Library/Application Support/dev.fella.app on macOS, and $XDG_DATA_HOME/dev.fella.app (usually ~/.local/share/dev.fella.app) on Linux. A custom XDG_DATA_HOME changes the Linux location. Deleting a local conversation cannot delete copies a provider may have retained under its own policies. Review those terms before sending sensitive data to a hosted model.

Untrusted file contents

Local documents and table cells can contain text that looks like instructions to the model. Fella labels workspace-derived content as untrusted evidence and instructs the model not to follow such text. The tool registry is fixed and read-only, so file contents cannot add a filesystem-write, shell, or web tool. This prompt boundary is defense in depth, not a guarantee that a model cannot be influenced by adversarial content; provider transmission still applies as described above. Web-page prompt-injection handling is not applicable until a web research tool is shipped.

Python runs inside a local capability boundary

run_python executes model-generated Python 3 inside an embedded wasm32-unknown-unknown RustPython guest under Wasmi. It has no filesystem, network, environment, clock, or subprocess access. The host provides captured stdout/stderr, OS entropy for interpreter startup, and a bounded read-only sql(query) bridge. It applies fuel, memory, stack, source, output, row, and response limits.
This is a stronger boundary for generated analytics, but it remains defense in depth: Wasmi, RustPython, and the checked-in guest artifact are trusted computing-base components. The guest does not include pandas, NumPy, SciPy, or a package installer.

Operational signals

Locally archived conversations include messages and evidence. The Settings Run Log is a local-only, content-minimized view of saved turn metrics and coarse friction signals; it does not duplicate prompts, answers, tool inputs/results, source names, or workspace paths. It is operational diagnostics, not a correctness score. The run log is derived from canonical turn records. Deleting a conversation removes unsaved associated records; analysis records retained to support saved outputs and reports remain with that saved work. Before using a hosted model with sensitive files, review that service’s retention, logging, and training terms.

Verification

See what Fella checks before presenting a result.

Experimental MCP

Read the closed extension boundary for future forks.