PaperGraph MCP

PaperGraph turns theorem-like environments in local or arXiv LaTeX papers
into a dependency graph exposed through MCP.

Development

uv sync
uv run pytest -q -p no:cacheprovider

Load a local paper

Call the MCP load_paper tool with the root .tex file. PaperGraph v0.2
recursively follows standard \input{...} and \include{...} commands.
Included paths are resolved relative to the file containing each command,
and .tex is optional in those commands.

The list_theorems, get_theorem, get_dependencies, and where_used
tools operate on the resulting combined graph.

Load an arXiv paper

PaperGraph v0.3 adds a dedicated load_arxiv_paper MCP tool:

load_arxiv_paper(
    arxiv_id="2401.12345",
    main_file=None,
    refresh=False,
)

It accepts modern identifiers (2401.12345, 2401.12345v2), legacy
identifiers (math/0307200, hep-th/9901001v3), and an optional arXiv:
prefix. URLs are not accepted. The tool downloads only from arXiv's fixed
https://export.arxiv.org/e-print/{id} source endpoint, so network access is
required for a paper that is not already cached.

Downloaded sources are stored in the user's platform cache. A later call for
the same identifier reuses the validated cache without another request. Pass
refresh=True to fetch a replacement; a failed refresh leaves the existing
valid cache intact.

PaperGraph normally finds the root file by looking for uncommented
\documentclass and \begin{document} commands, preferring main.tex,
paper.tex, and manuscript.tex when needed. If multiple candidates remain,
the error lists them. Retry with a project-relative path such as
main_file="src/article.tex" to choose one explicitly.

For safety, PaperGraph streams downloads and extraction, limits a compressed
response to 100 MiB, limits extracted content to 500 MiB and 10,000 members,
and rejects path traversal, links, devices, FIFOs, and other special archive
members. Tar archives, compressed tar archives, gzip-compressed single TeX
files, and plain single-file TeX responses are supported.