PaperGraph MCP

Community

Read math papers with evidence, not guesses.

How to install ↓ Compare GitHub source repository ↗

Category: Other #106 of 230
Stars: ★ 279 Very popular
Transport: stdio (local process)
Runtime: Python 3.10+ · Prebuilt binary uv (uvx)
Credentials: No credential needed
License: MIT
Last commit: 14 days ago
Tools: 54

55FMRS · C

PaperGraph MCP is an evidence-first reading server for math papers, published from a personal repository under the MIT license and speaking MCP over stdio. Its core value is tying results, proofs, and citations to source spans that can be sliced back out of the original paper, while explicitly framing empty results as extraction limits rather than mathematical facts. It suits readers who insist on reviewing evidence, and does not suit those expecting proof verification, semantic equivalence judgments, or automatic crawling of the citation chain.

Strongest · Documentation 14/20 Weakest · Reliability 8/20

Reliability

8/20

Security and permissions

12/20

Maintenance

10/20

Documentation

14/20

Setup experience

11/20

Why each score

Reliability 8/20: Static review relies only on the README and description: it claims a stable v1.1.3, shows a CI badge and a pytest suite, but supplies no verifiable source, test files, or CI workflow content, so this is self-reported rather than execution evidence, and the static calibration caps reliability at 12. The docs do disclose evidence boundaries (empty dependency results mean extraction limits, not mathematical facts), weak scanned-PDF handling, and that the parser is not a full TeX engine, which is a plus. However, tool behavior cannot be checked against the declared index, and error handling is only described in general terms, so the score is reduced.
Security and permissions 12/20: No malware, credential theft, covert exfiltration, or unconfirmed destructive defaults were found; the README states downloads are built only from arXiv's fixed e-print endpoint, arbitrary URLs are rejected, and archives are bounded at 100 MiB compressed / 500 MiB expanded / 10000 members with rejection of absolute paths, parent traversal, symlinks and special members, which are positive controls. Deductions: external metadata search contacts Crossref, OpenAlex and arXiv with only broad data-flow disclosure; user confirmation for side-effecting import/parse operations is not fully specified; least-privilege and credential handling are not detailed. Because none of this is verifiable against source and confirmation/scoping is incomplete, a mid-to-upper score is given.
Maintenance 10/20: There is an MIT license, semantic versioning, and a fairly detailed release history from v0.4.0 to v1.1.3 indicating continued iteration, plus 196 stars and only 1 open issue. However, the static material contains no commit timeline, issue-response record, dependency-update policy, or security-response channel (e.g., SECURITY.md), leaving governance and maintenance evidence insufficient to conclude sustained maintenance, hence a mid-range score.
Documentation 14/20: The documentation is layered and bilingual (EN/CN) with quick start, MCP client JSON config examples, a full tool reference, CLI examples, evidence-boundary notes, safety/privacy limits, disclosed limitations (scanned PDFs, explicit main_file for complex projects, not a full TeX engine), and reviewable example artifact paths, which is above average. Deductions: some details require following external docs, the manifest is not cached, and full install/permission/cost specifics still depend on repository files not provided here, so it does not reach 20.
Setup experience 11/20: The install path is clear: uvx pinned to the v1.1.3 tag, a standard MCP client stdio JSON config, a doctor self-check, and a first-workspace walkthrough, with few steps. But static review cannot confirm the install actually works, it requires the user to install uv, edit client config and restart the client, and examples are mostly PowerShell with cross-platform variation not fully covered. Under static calibration setup cannot exceed 15, and the absence of verifiable execution evidence lowers it further.

Static review · not runListed 2026-09-18

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local filesUses the networkConnects to a database

Best for

Math-paper readers who need a strict separation between evidence and guesses
Researchers who want an AI agent to help map theorems and dependencies inside a local workspace
People who want reading artifacts as Markdown they can keep in Git or notes
Users who want to review external references before importing them

Not for

People expecting the tool to prove or verify mathematical correctness
Anyone wanting automatic infinite crawling of the citation chain (A cites B, B cites C)
Users needing semantic theorem matching or equivalence judgments for similarly worded results
Users with scanned or OCR-heavy PDFs where extraction may be sparse
Users who expect a full TeX engine for complex projects

Required permissions

Read and write SQLite workspace files at user-chosen local paths
Read user-supplied local LaTeX projects and local PDF files
Download e-prints from arXiv's fixed e-print endpoint
Query public metadata services (Crossref, OpenAlex, arXiv) when the user runs a reference search
Write generated Markdown reports and manifest files into a user-specified artifact directory

Risks and side effects

The workspace stores text, source spans, and proof evidence extracted from papers, so committing databases or private manuscripts can leak content
The Scholarly Reference Resolver only searches public metadata and stops with explicit boundaries for ambiguous, old, paywalled, or metadata-only literature, requiring manual sourcing
An empty dependency result only means no resolvable theorem-label references were found under the current rule; it is not evidence that the theorem has no mathematical dependencies
Scanned PDFs may yield sparse text and missing evidence
The parser is not a full TeX engine, so complex projects may need an explicit main_file

Setup

Before you start

Runtime：Python 3.10+ · Prebuilt binary uv (uvx)

Install uv (see the uv installation docs).
Verify the pinned release without cloning: uvx --from git+https://github.com/lotchuazzz-crypto/[email protected] papergraph-mcp --version and ... papergraph-mcp doctor.
Add a mcpServers.papergraph entry to an MCP client that accepts JSON-style stdio configuration (command uvx, args pinned to the v1.1.3 tag).
Restart the MCP client after changing its configuration.
The server uses stdio, so running it without --help or --version waits quietly for an MCP client connection.
If you already have a clone, run git fetch --tags origin first so the checkout is not stuck on an old local origin/main.

MCP config · JSON

{
  "mcpServers": {
    "papergraph": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/lotchuazzz-crypto/[email protected]", "papergraph-mcp"]
    }
  }
}

Works in clients that read an "mcpServers" key. For clients such as VS Code that use a different key, the configurator below converts it.

.vscode/mcp.json

{
  "servers": {
    "papergraph": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/lotchuazzz-crypto/[email protected]",
        "papergraph-mcp"
      ]
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal

claude mcp add papergraph -- uvx --from git+https://github.com/lotchuazzz-crypto/[email protected] papergraph-mcp

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

After adding the config and restarting the MCP client, tools such as workspace_get_paper_map and load_arxiv_request should appear in the client's tool list; you can also run `uvx --from git+https://github.com/lotchuazzz-crypto/[email protected] papergraph-mcp doctor` first to confirm the install.

Troubleshooting

Run papergraph-mcp doctor or get_environment_diagnostics to check the environment
Confirm you pinned the v1.1.3 tag so installations stay reproducible
No output when running the command directly is expected: the server uses stdio and waits for an MCP client connection
Restart the MCP client after editing its configuration
For an existing checkout, run git fetch --tags origin before treating it as current
If validate_arxiv_request returns action: ask_user_to_choose, ask the user to choose rather than continuing
When extraction looks sparse, check workspace_get_dependency_diagnostics and Evidence Triage to see why

Things to try

Once connected, you can ask your AI assistant things like:

Load the arXiv paper math/0307200 and show me its Paper Map.
Add this local PDF paper to the workspace and list all theorem-like results.
Create a reading queue for local:paper-a::thm:main and export a single-paper Reading Report.
Search metadata candidates for this blocked external reference, then import the candidate I choose.

Tools 54

open_workspace writes

Open a local SQLite workspace

workspace_add_local_paper writes

Add a local LaTeX paper to the workspace

workspace_add_arxiv_paper writes

Add a paper by arXiv ID to the workspace

workspace_add_pdf_paper writes

Add a PDF paper to the workspace

workspace_list_papers read-only

List papers in the workspace

workspace_get_paper read-only

Get details for one paper

workspace_search_theorems read-only

Search theorem-like results in the workspace

workspace_get_dependencies read-only

Get dependencies of a result

Show 46 more tools

workspace_get_dependency_diagnostics read-only

Get diagnostics for dependency extraction

workspace_get_citations read-only

Get citation relations, optionally including unresolved ones

workspace_get_paper_map read-only

Build a Paper Map: main-result candidates, structure, and external reading risks

workspace_export_paper_reading_report read-only

Export a single-paper Markdown reading report

workspace_export_cross_paper_reading_plan read-only

Export a cross-paper Markdown reading plan

workspace_plan_starter_project read-only

Plan a starter project producing START_HERE.md and a starter manifest

workspace_bootstrap_reading_project writes

Bootstrap reading-project artifacts from explicit paper inputs

workspace_list_results read-only

List extracted theorem-like results

workspace_get_result read-only

Get one extracted result

workspace_get_result_proof read-only

Get the proof evidence for a result

workspace_get_proof_dependencies read-only

Get dependency evidence inside proofs

workspace_get_external_result_mentions read-only

Get mentions of external results

workspace_get_evidence read-only

Get evidence with source spans

workspace_export_reading_bundle read-only

Export a reading bundle of evidence

workspace_export_result_reading_context read-only

Export the reading context for a result

workspace_get_source_slice read-only

Slice the corresponding text back out of the original paper

workspace_get_result_reading_path read-only

Get the reading path for a result

workspace_create_reading_session writes

Create a reading session

workspace_list_reading_sessions read-only

List reading sessions

workspace_get_reading_session read-only

Get a reading session

workspace_record_reading_checkpoint writes

Record a reading checkpoint

workspace_add_reading_note writes

Add a note to a reading session

workspace_export_reading_session_summary read-only

Export a reading session summary

workspace_create_reading_queue writes

Generate a reading queue from local proof evidence

workspace_list_reading_queues read-only

List reading queues

workspace_get_reading_queue read-only

Get a reading queue

workspace_apply_reading_queue_to_session writes

Apply a reading queue to a reading session

workspace_plan_external_imports_for_result read-only

Plan external imports for a result

workspace_plan_external_imports_for_queue read-only

Plan external imports for a reading queue

workspace_plan_external_imports_for_paper read-only

Plan external imports for a paper

workspace_search_external_reference read-only

Search Crossref, OpenAlex, and arXiv metadata for a blocked reference

workspace_list_external_reference_searches read-only

List external reference searches

workspace_resolve_external_reference_candidate writes

Apply a chosen search candidate through the reference closure workflow

workspace_resolve_external_reference writes

Resolve a blocked reference via DOI, URL, published metadata, or local PDF

workspace_list_external_reference_resolutions read-only

List external reference resolutions

get_environment_diagnostics read-only

Return environment diagnostics

validate_arxiv_request read-only

Validate arXiv input and report the decision without loading

load_arxiv_request writes

High-level entry point that validates and loads an arXiv paper

validate_arxiv_input read-only

Validate arXiv input

load_paper writes

Load a local paper

load_arxiv_paper writes

Load a paper by an already-disambiguated arXiv ID

list_theorems read-only

List theorem-like results in a paper

get_theorem read-only

Get a single theorem

get_dependencies read-only

Get theorem dependencies

get_dependency_diagnostics read-only

Get dependency diagnostics

where_used read-only

Find where a result is used

Use cases

Load an arXiv paper, a local LaTeX project, or a born-digital PDF and produce a Paper Map

Inspect theorem statements, proof evidence, source slices, and dependency diagnostics

Generate a reading queue from local proof evidence and keep reading sessions and checkpoints

Export a deterministic Markdown Reading Report as a durable reading handoff

Export a Cross-Paper Reading Plan for a few explicitly given related papers

Search public metadata for blocked references and close the loop only on a user-chosen candidate

Supported clients

MCP clients accepting JSON-style stdio configuration

Listed from the project's documentation, not tested by this site.

Overview

PaperGraph MCP is a local, evidence-first reading workflow server for AI agents working with math papers. It turns arXiv papers, local LaTeX projects, and born-digital PDFs into a theorem-centered local workspace so a researcher can inspect where every claim came from: extract results, trace proof evidence, plan reading order, review external dependencies, and keep reading state locally. It communicates over stdio, stores workspaces as ordinary local SQLite files, and adds a Scholarly Reference Resolver in v1.1.3 for metadata search of blocked references across Crossref, OpenAlex, and arXiv.

Similar servers

Hevy MCP 77 · B

Manage your Hevy workout data from AI assistants via MCP.

★ 529 · Tools 26 Compare with this →

Firecrawl MCP Server 75 · B

Firecrawl's official MCP server for web search, scraping, and structured extraction for AI agents

★ 7.7k · Tools 11 Compare with this →

BioMCP 75 · B

One binary. One grammar. Evidence from the biomedical sources you already trust.

★ 670 · Tools 11 Compare with this →

Compartment 75 · B

Encrypted, fully offline long-term memory for AI agents

★ 579 · Tools 14 Compare with this →

Compare with other MCP servers

Source revision af7f30280beb Data synced 2026-11-02 Read the FMRS scoring method