**doc7** is a Go command-line tool that converts documents into AI-ready Markdown by routing every page through a **multimodal vision model** instead of a traditional OCR stack. PDFs, Office files, scans, screenshots, charts, formulas, and diagrams all enter the same page-understanding pipeline and leave as one searchable Markdown document, with the ordered pages rebuilt into a single output. The architectural bet is that character extraction is the wrong primitive for modern documents. Conventional converters read a text layer and fall back to OCR when one is missing, which loses figure semantics, equation structure, and the ordered relationships inside diagrams. doc7 instead asks a vision-language model to read the whole page. Its published benchmark makes the claim concrete: on two raster-only PDFs scored against **15 machine-checkable visual facts**, doc7 recovered **15/15**, compared with 9/15 for MarkItDown with its OCR plugin and 3/15 for Docling's standard local pipeline. doc7 and MarkItDown both used the same `qwen3.5-9b` model through the same local OpenAI-compatible endpoint, so the gap reflects pipeline design rather than model choice. A worked example runs a raster-only page from *Attention Is All You Need* and recovers Figure 2, the displayed equation, the scaling rationale, and the diagram relationships. Crucially, doc7 supplies no model of its own. It targets any **OpenAI-compatible multimodal endpoint** — LM Studio or Ollama on the same machine, a private deployment, or a remote provider. First run auto-discovers running local model servers, reads their real model IDs, verifies image understanding, and saves the selection. That means no document-processing service lock-in and no per-page parser fee, though it also means output quality is bounded by whichever vision model the user supplies. One binary carries the interactive CLI, batch processing, model checks, an MCP server, a Go SDK, and an asynchronous HTTP service, so the same tool serves both a person converting a report and an agent pipeline ingesting a corpus. The project is young and moving quickly — created 2026-08-02 under the MIT license, with v0.1.2 tagged on 2026-08-07. macOS code signing and notarization are noted as planned rather than shipped, so macOS users currently clear the download quarantine attribute manually or install via the checksum-verified installer script.