A document parser that finally looks at the page

What it does

doc7 is a Go CLI that ingests PDFs, Office files, scans, and images and emits structured Markdown. Instead of scraping text layers or chaining traditional OCR tools, it renders each page as an image and passes it to an OpenAI-compatible multimodal model for interpretation. The output is a single searchable document that preserves text, tables, formulas, charts, and diagram relationships. The same binary also provides an interactive chat agent, a Go SDK, an asynchronous HTTP service, and an MCP interface.

The interesting bit

The project treats document conversion as a computer-vision problem rather than an extraction pipeline. By pushing rasterized pages through a single local quantized VLM, it sidesteps the usual brittle stack of separate layout, table, and formula parsers. The README cites a specific benchmark run on a raster-only academic paper where doc7 recovered fifteen of fifteen checked visual facts, compared with nine for MarkItDown OCR and three for Docling using their standard local pipelines in that same test.

Key highlights

• Understands text, tables, formulas, charts, and diagrams through one unified visual pipeline instead of specialized single-purpose parsers.
• Runs fully offline against local endpoints such as LM Studio or Ollama, with no per-page or per-conversion metering from the tool itself.
• Bundles a chat agent that uses OpenAI-compatible Tool Calling to navigate files and invoke conversion without exposing arbitrary shell access.
• Warns explicitly before sending content to remote endpoints, and stores any required API keys in the OS keychain or credential manager rather than config files.
• Distributed as a single binary for macOS, Linux, and Windows with a built-in updater that verifies release checksums.

Caveats

• macOS releases currently lack Apple Developer ID signing and notarization; the README notes you must manually strip the quarantine attribute after extracting the archive.
• The published benchmark is a single measured run on one paper, and the README itself notes that quality and speed depend heavily on which model and quantization you choose.
• Models that do not support Tool Calling can still chat, but they lose the restricted filesystem browsing and guided conversion features.

Verdict

Grab it if you already run local vision models and want to keep document processing private and off metered cloud APIs. Pass if you need deterministic, lightning-fast extraction or cannot accommodate the memory and compute footprint of a multimodal model.