PaperDebugger (MCP server) — Use Cases, Install & Live Demo

Name: PaperDebugger MCP Server
Author: PaperDebugger

Why use it

Key features

Multi-agent (Reviewer, Editor, Citation Auditor, Coherence)
Plugin-based — write your own agent
Works on LaTeX, Markdown, and PDF input
Citation hallucination guard (cross-checks against bib + Semantic Scholar)
In-editor diffs, not raw rewrites
Open source

Live Demo

What it looks like in practice

paperdebugger-mcp.replay ▶ ready

0/0

Install

Pick your client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "paperdebugger-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "paperdebugger-mcp"
      ]
    }
  }
}

Open Claude Desktop → Settings → Developer → Edit Config. Restart after saving.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "paperdebugger-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "paperdebugger-mcp"
      ]
    }
  }
}

Cursor uses the same mcpServers schema as Claude Desktop. Project config wins over global.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "paperdebugger-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "paperdebugger-mcp"
      ]
    }
  }
}

Click the MCP Servers icon in the Cline sidebar, then "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "paperdebugger-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "paperdebugger-mcp"
      ]
    }
  }
}

Same shape as Claude Desktop. Restart Windsurf to pick up changes.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "paperdebugger-mcp",
      "command": "npx",
      "args": [
        "-y",
        "paperdebugger-mcp"
      ]
    }
  ]
}

Continue uses an array of server objects rather than a map.

~/.config/zed/settings.json

{
  "context_servers": {
    "paperdebugger-mcp": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "paperdebugger-mcp"
        ]
      }
    }
  }
}

Add to context_servers. Zed hot-reloads on save.

claude mcp add paperdebugger-mcp -- npx -y paperdebugger-mcp

One-liner. Verify with claude mcp list. Remove with claude mcp remove.

Use Cases

Real-world ways to use PaperDebugger

How to get a tough fake-peer-review pass before submitting

👤 PhD students, researchers, paper authors ⏱ ~60 min intermediate

When to use: You're 24h from submission and want one more brutal review.

Prerequisites

LaTeX source or compiled PDF — Either format works

Flow

Run Reviewer

Use PaperDebugger Reviewer on /paper/main.tex. Be harsh — venue is NeurIPS-style, top-tier.✓ Copied

→ Numbered review with strengths, weaknesses, suggested experiments
Citation audit

Cross-check every \cite. Anything fabricated, broken, or missing from the bib?✓ Copied

→ Per-citation status table; flagged entries
Coherence

Flag claim drift between abstract / introduction / conclusion. Quote the conflicting sentences.✓ Copied

→ Drift report with quoted pairs

Outcome: A pre-submit checklist resolved before reviewer 2 ever sees the paper.

Pitfalls

Reviewer agent over-suggests experiments you can't run — Filter the suggestions list — keep only those that improve the writing without new data

Combine with: filesystem

Polish the language without losing the author's voice

👤 Non-native English authors ⏱ ~45 min beginner

When to use: The science is fine; the prose has rough patches.

Flow

Editor pass

Use the Editor agent on /paper/main.tex. Preserve voice; only fix grammar, awkward phrasing, and ambiguous antecedents.✓ Copied

→ Diff list, sentence by sentence
Review diffs

Show me the 10 highest-impact edits — ones that fix actual ambiguity, not stylistic preferences.✓ Copied

→ Top 10 with rationale

Outcome: Crisper paper, same voice.

Pitfalls

Editor 'corrects' technical jargon — Pass --preserve-glossary terms.txt so it leaves listed terms alone

Verify every result table matches what's in the experiments directory

👤 ML researchers paranoid about table errors ⏱ ~90 min advanced

When to use: You ran results 3 different ways; not sure which numbers ended up in the paper.

Prerequisites

Paper source + experiments directory with raw results (JSON/CSV) — Standard folder layout

Flow

Map tables to results

Map every numerical claim in /paper/main.tex to a specific file in /experiments/. Flag claims you can't trace.✓ Copied

→ Mapping table; un-traced claims listed
Verify

For each traced claim, recompute from the raw file and report match/mismatch.✓ Copied

→ Verification table with deltas

Outcome: Confidence that the numbers in the paper match the experiments — or a list of fixes.

Pitfalls

Experiments folder is ad-hoc — no convention — Provide a hint file mapping table labels to experiment paths

Combine with: filesystem

Combinations

Pair with other MCPs for X10 leverage

paperdebugger-mcp + filesystem

Operate on a working copy and persist the diffs

Save the editor's diffs as /paper/edits-$(date +%F).patch.✓ Copied

paperdebugger-mcp + arxiv-mcp-server

Pull related work to feed the Reviewer agent

Fetch top 10 arxiv papers cited in the related work, then re-run Reviewer with that context.✓ Copied

Tools

What this MCP exposes

Tool	Inputs	When to call	Cost
review	paper_path, venue?	Pre-submission review	Heavy LLM call
edit	paper_path, preserve_glossary_path?	Language polish	LLM
audit_citations	paper_path, bib_path	Always before submission	1 Semantic Scholar lookup per citation
check_coherence	paper_path	Late-stage editing	LLM

Cost & Limits

What this costs to run

API quota: Bound by your underlying LLM
Tokens per call: 5000–25000 per agent run
Monetary: Free; LLM cost is yours
Tip: Run agents serially (Editor first, Reviewer second) and pass diffs forward to avoid re-tokenizing the whole paper

Security

Permissions, secrets, blast radius

Credential storage: LLM API key via env

Data egress: Your LLM provider + Semantic Scholar (for citation audit)

Anonymous-submission venues: don't include identifying paths in prompts unless your LLM provider has zero-retention

Troubleshooting

Common errors and fixes

Editor diffs break LaTeX compile

Run with --syntax-aware so it preserves \macro{...} braces and math

Verify: Compile after applying diffs

Citation audit flags real citations as fabricated

Citation may be too new for Semantic Scholar — pass it via --known-good citations.txt

Reviewer is too generic

Pass --venue NeurIPS (or specific) so style and standard match

Alternatives

PaperDebugger vs others

Alternative	When to use it instead	Tradeoff
Writefull / Grammarly	You only need language polish, not multi-agent review	Less powerful; doesn't audit citations or coherence
Roll-your-own Claude prompt	You only review one paper a year	PaperDebugger has shaped prompts and a coordination loop you'd otherwise rebuild

More

Resources

📖 Read the official README on GitHub

🐙 Browse open issues

🔍 Browse all 400+ MCP servers and Skills