Turn any document into clean, AI-ready Markdown. Local-first desktop app: reads scanned PDFs, batches folders, runs offline, and uses far fewer tokens than vision models.
Python desktop-appdocument-conversionllmlocal-firstmarkdownmarkitdownocrofflinepdf-to-markdownragrustsveltetauri
Changelog
All notable changes to this project are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
[Unreleased]
[0.3.1] - Unreleased
Fixed
- Lite editions could not complete first-run setup on Windows and Linux: the app downloaded
uv 0.5.11 while the shipped lockfiles are compiled with uv 0.12 (
--require-hashesfailed with "all requirements must be pinned upfront ... found: tzdata"). The app now installs the same uv 0.12.3 that CI validates the lockfiles with, and CI fails if the two ever drift (#57). - Linux portable archives failed to start with "requirements.lock not found at /usr/lib/MDFlux/..." unless they were symlinked into the system location. Resources are now resolved next to the executable, so the archive runs from any directory (#53).
- Diagnostics reported HTTP 403 for valid Groq keys because the probe sent Python's default
user agent. Every provider request now identifies as
MDFlux/<version>(#25). - Lite installs a versioned uv into
app-data/bin/uv-<version>/through a staging directory, so a half-downloaded or obsolete installer is never reused.
Changed
- Release and release-candidate workflows launch each packaged archive in an isolated
package-verification mode (
--app-smoke): first run, restart, and relocation of the payload must pass before an archive can be published. Expected archive names are derived from the application version instead of being hard-coded.
0.3.0 - 2026-08-23
Added
- Cross-platform Windows x64 and Linux x64 (glibc) release packaging with Lite and Full editions.
- Linux platform support, including Ubuntu/glibc baseline guidance and WebKitGTK requirements.
- Cross-platform documentation covering editions, supported platforms, Ubuntu contributor setup, release archive names, checksums, and artifact verification.
- Bug report fields for platform, edition, architecture, and runtime status.
Changed
- Lite now uses transactional, recoverable runtime provisioning; Full bundles an immutable offline-ready runtime and never installs Python packages at runtime.
- Runtime health, repair, and edition diagnostics now make provisioning state and failures clearer.
- AI cleanup handling and provider/runtime diagnostics were cleaned up for more predictable runs.
- README and contributor guidance now describe one shared Windows and Linux product codebase.
Fixed
- Addressed AI cleanup timeout resilience reported in #25, Linux support reported in #28, and OCR dependency/provisioning failures reported in #30 through the bundled Full runtime and recoverable Lite provisioning.
0.2.0 - 2026-08-19
Added
- Changes tab after cleanup, with added and removed lines.
- Provider picker in Diagnostics (DeepSeek, OpenAI, Groq, OpenRouter, or a compatible endpoint). Test the key first. OpenAI and DeepSeek keys both start with
sk-, so pick those yourself; other prefixes switch automatically. - Drop
.txtand.mdlike any other file. - Banner when a file converted to nothing usable.
- Type chip when dragging a file over the window.
- Ctrl+C with nothing selected copies the full Markdown.
- Clearer errors for damaged PDFs, unsupported types, and empty extracts.
Changed
- Better conversion for PowerPoint charts, SVG, and Word equations.
- Better on-device OCR for scans. Models still ship in the install.
- JSON and XML show as fenced code in Preview.
If you already had OCR from 0.1.0, reinstall it from Diagnostics.
0.1.0 - 2026-06-19
First public release. Windows, portable (extract-and-run; no installer).
Added
- Convert PDF, DOCX, PPTX, XLSX, EPUB, HTML, CSV, JSON, and XML to clean Markdown, built on Microsoft's MarkItDown.
- Cleanup modes: Off, deterministic rule-based, and an optional AI pass (local Ollama or bring-your-own-key OpenAI-compatible / Anthropic).
- OCR for scanned PDFs and images (RapidOCR) and audio transcription (faster-whisper), installed on demand as optional engines.
- Batch conversion with adaptive concurrency, cancel, timeouts, and per-file progress.
- Output control: folder rules, naming templates, before/after preview.
- Self-provisioning Python 3.12 runtime on first launch; fully offline thereafter.
- First-run setup shown as a multi-step stepper with live download size/speed.
Security
- Dependencies are integrity-verified during the one-time setup, so first run is trustworthy. See
SECURITY.md.