Optical Context MCP

It reads a local PDF, runs OCR with Mistral, recomposes the extracted text and figures into dense PNGs, and exposes those artifacts over MCP for batch retrieval.

What It Does

• reads a local PDF from the MCP host machine
• extracts page markdown and embedded images with Mistral OCR
• packs that content into dense PNGs that preserve visual grouping
• optionally sizes embedded figures with a bundled technical-document model
• stores a manifest and temp job artifacts for follow-up retrieval
• lets an agent pull only the packed images it needs

Where It Fits

Use it for:

• operating manuals
• scanned handbooks
• product catalogs
• PDF slide decks
• visually structured OCR-heavy documents

Skip it for:

• tiny PDFs
• clean text-native PDFs where normal extraction is enough
• workflows that require exact page-faithful rendering
• cases where OCR cost is not justified

Example Result

The image below shows a real local validation run on a public research paper with dense text, figures, charts, and page-level visual structure. The packed image on the right consolidates the seven source pages shown on the left. Example local run facts from the generated manifest:

• source paper pages: 22
• previewed source page range: 15 to 21
• extracted images: 30
• packed output images: 6
• example packed image size: 986x1084
• example packed image file size: 536,697 bytes

This example shows the intended workflow: take a long, visually structured PDF and compress it into a smaller set of retrievable packed images that still preserve the visual structure of the source.

Install

python -m pip install optical-context-mcp

Install with the adaptive sizing runtime:

python -m pip install "optical-context-mcp[ml]"

Run without installing:

uvx optical-context-mcp

• MISTRAL_API_KEY is required for compress_pdf
• packed images are always stored locally under the system temp directory
• compress_pdf returns up to 30 packed images inline by default
• the adaptive sizing checkpoint is bundled with the package
• adaptive sizing activates automatically when torch and torchvision are available
• set OPTICAL_CONTEXT_DISABLE_ADAPTIVE_SIZING=1 to force the legacy fixed sizing
• set OPTICAL_CONTEXT_ADAPTIVE_MODEL_PATH=/path/to/model.pt to override the bundled checkpoint

For pinned shared setups:

uvx --from optical-context-mcp==0.1.5 optical-context-mcp

Run

Default transport is stdio:

optical-context-mcp

Claude Code

Register the server in a project:

claude mcp add -s project optical-context -- uvx optical-context-mcp

Typical use:

1. call compress_pdf
2. inspect the returned manifest
3. fetch packed images with get_packed_images

MCP Tools

• compress_pdf: run OCR plus recomposition and create a stored job
• get_job_manifest: load metadata for an existing job
• get_packed_images: fetch one or more packed PNGs from an existing job

How It Works

flowchart LR
    A["Local PDF"] --> B["Mistral OCR"]
    B --> C["Page markdown + embedded images"]
    C --> D["Recomposition engine"]
    D --> E["Dense packed PNG images"]
    E --> F["Stored job artifacts"]
    F --> G["Agent fetches manifest or image batches over MCP"]

Why Packed Images Instead Of Just OCR Text

• section grouping
• table-like layout
• captions near figures
• visual adjacency between text and embedded graphics

For many vision-capable agents, that is a better intermediate format than a plain OCR dump.

Current Scope

• depends on Mistral OCR
• currently handles local file paths, not remote uploads
• stores artifacts in the local system temp directory by default
• optimized for compression and retrieval, not final polished markdown generation
• quality depends on OCR quality and the visual density of the source document
• adaptive sizing falls back safely to fixed medium image sizing when the ML runtime is absent

Roadmap

• make the OCR layer provider-agnostic so different OCR backends can be swapped behind the same MCP workflow

Feedback

Tried Optical Context in a document workflow? You can voluntarily share what worked and what got in the way using the feedback form. Include your package version, MCP client and a brief description of the document type. Do not attach private PDFs or API keys. Feedback is public on GitHub.

See the changelog for changes between published versions.

Development

uv venv --python /opt/homebrew/bin/python3.11 .venv
uv pip install --python .venv/bin/python -e ".[dev]"
.venv/bin/python -m pytest

Publishing

Keep pyproject.toml and both versions in server.json synchronized. Run the Publish PyPI workflow on the reviewed release commit. It checks the wheel, runs tests against the installed distribution, publishes to PyPI, then publishes the same version to the MCP Registry. Both use GitHub OIDC; no local publishing token is required. If only the registry job fails, rerun that failed job rather than republishing the immutable PyPI version. When a workflow correction requires a new run, select registry_only to retry registry publication without uploading the existing PyPI version again.

Create the matching GitHub release from that same commit and attach the workflow's python-distributions artifact.