Papergraph MCP
PaperGraph MCP turns math papers into evidence-grounded reading maps for AI agents: extract results, trace proof evidence, plan reading order, and review external dependencies without guessing.
Papergraph MCP is an open-source mcp server by lotchuazzz-crypto, written mainly in Python, with 279 GitHub stars and the MIT license. Its code was last updated Oct 18, 2026. The latest PyPI release is papergraph-mcp 1.2.3, published Oct 18, 2026, with 39 downloads last week.
PaperGraph MCP
Read math papers with evidence, not guesses.
Turn arXiv papers, local LaTeX projects, and born-digital PDFs into a local theorem-centered workspace. Inspect source-backed results, proofs, citations, and dependencies; save reading progress and export Markdown reports.
</div>What You Can Do
- Start with a Paper Map: inspect structure, main-result candidates, and external reading risks before choosing a theorem.
- Trace statements and proof-local references back to source slices; inspect direct and recursive dependencies, shared prerequisites, cycles, and unknowns.
- Discover public full-text candidates for DOI papers and review scholarly reference matches from Crossref, OpenAlex, and arXiv.
- Keep queues, sessions, checkpoints, notes, and open questions in local SQLite workspaces. Export Reading Reports and Cross-Paper Reading Plans.
Quick Start
Install uv; no Git or checkout is required:
uvx --from papergraph-mcp==1.2.3 papergraph-mcp --version
uvx --from papergraph-mcp==1.2.3 papergraph-mcp doctor
PyPI 1.2.3 is published and has passed actual public-package and local Git-free acceptance. The dedicated GitHub public-index CI has not run; the GitHub Release is not published. Diagnostics alone do not verify publication.
An MCP client can use this JSON-style stdio configuration:
{
"mcpServers": {
"papergraph": {
"command": "uvx",
"args": ["--from", "papergraph-mcp==1.2.3", "papergraph-mcp"]
}
}
}
Restart the client after configuration. A bare server launch waits quietly for a stdio MCP connection; CLI commands also work without an MCP client.
For agent-assisted setup, use the repository's setup skill from the freshly fetched remote default revision. Preserve existing branches and local changes; ask before installing software, changing client configuration, or restarting. The optional Git-source route requires Git and live verification of the selected version tag against its reviewed package source.
Read Your First Paper
- Keep the workspace database outside the repository. Preview inputs with
plan-starter-project; it validates files without importing or writing. Runbootstrap-reading-projectafter review to create the workspace and reading artifacts. Local LaTeX inputs must name the main.texfile. - Open the Paper Map and Evidence Triage, choose a result, then inspect its statement, proof, source slices, and dependency evidence.
- Review external import plans, save checkpoints and notes, and export a report or cross-paper plan. Follow the first-workspace walkthrough for commands and examples.
For raw arXiv requests, use load_arxiv_request; conflicting IDs or links require user selection before loading. For DOI discovery, discover-doi DOI finds candidates without downloading. add-doi-paper WORKSPACE DOI can import a requested root paper when exactly one eligible public PDF exists; multiple candidates require selection.
For an existing workspace, replace the placeholders with actual IDs:
uvx --from papergraph-mcp==1.2.3 papergraph-mcp get-paper-map --workspace WORKSPACE --paper-id PAPER_ID
uvx --from papergraph-mcp==1.2.3 papergraph-mcp get-dependency-reading WORKSPACE PAPER_ID --target-result-id RESULT_ID --detail compact
Arguments are command-specific: check each command's --help. Compact dependency output retains access to full evidence.
Evidence Boundaries
- PaperGraph does not verify proofs, infer hidden prerequisites, or establish semantic theorem equivalence. Empty dependencies do not prove independence; extracted reading orders are not complete mathematical dependency graphs.
- Long PDFs support raw source pagination through
workspace_read_source_blocksand itsnexthandle. New automatic PDF proof evidence covers only the first physical line with an explicit proof heading; ownership can remain unknown, and later proof references remain unknown. Raw continuation is readable but does not become dependency evidence automatically. Existing excerpts retain their extraction policy until explicit reimport. - External references become reviewable plans, not automatic downloads. Bounded reference expansion requires an explicitly approved finite policy; ambiguous matches await review. Metadata is not full text, and no workflow bypasses paywalls.
- Papers and reading state remain local. Back up workspaces before schema upgrades; older versions cannot open schema 9.
Documentation
- v1.2.3 release notes · Publication evidence · Artifact verification
- Local LaTeX extraction and independent acceptance
- Core contract · Historical v1.1.5 client evidence
- Examples: Reading Report · Cross-Paper Reading Plan · Reference Search · Reference Resolution
- Starter outputs: Summary · Manifest
- v1.2.2 release notes · Publication evidence · Artifact verification
- Contributing · MIT License
Papergraph MCP alternatives
View more →Similar mcp servers in Research & Analysis, ranked by GitHub stars. Compare licence, recent activity and scan status before you choose.
| Name | Stars | Licence | Last updated | Scan |
|---|---|---|---|---|
| Papergraph MCP (this page) lotchuazzz-crypto | 279 | MIT | Oct 18, 2026 | Scanned, no flagged issues |
| Worldmonitor koala73 | 90k | AGPL-3.0 | Nov 1, 2026 | Scanned, no flagged issues |
| GPT Researcher assafelovic | 30k | Apache-2.0 | Oct 10, 2026 | Scanned, no flagged issues |
| Auto Claude Code Research In Sleep wanshuiyin | 18k | MIT | Nov 1, 2026 | Scanned, no flagged issues |
| Deep Research u14app | 4.7k | MIT | Jun 18, 2026 | Scanned, no flagged issues |
| MCP Server Chart antvis | 4.4k | MIT | Aug 27, 2026 | Scanned, no flagged issues |
