latex-mcp-server

LaTeX MCP Server

An MCP (Model Context Protocol) server that can supercharge your VSCode + LaTex Workshop writing workflow

Currently houses MCP tools, and new tools can be easily added in either Python and TypeScript. The Python process is the MCP server; selected tools are delegated to a small Node/TypeScript module via a bridge.

LaTex MCP Server + Copilot/Claude/Cursor, what can you do?

• Read your cited papers, verify / add onto your claims
• Access your python scripts to generate new figures / latex tables based on your experiment data, and place them into latex autonomously
• Can verify every step and ensure your pdf compiles
• Functionally it can autonomously take your outline + experimental results and turn it into a paper (may looks like slop though)

<p align="center"> <video src="./docs/demo/minimal.mp4" controls width="600"> Your browser does not support the video tag. </video> </p>

---

Python Tools

• list_tex_files

Lists all .tex files under your LaTeX workspace (relative paths). Useful for LLMs or agents to discover available chapters and sections.

• read_file

Reads a text or binary-safe slice of a file given a relative path inside the workspace. Supports specifying a maximum number of bytes to return.

• extract_bibliography

Parses a BibTeX file and returns structured entry metadata, including download URLs (DOI, arXiv). Helps agents or scripts work with citation data programmatically.

• download_bibliography

Downloads PDFs for bibliography entries listed in resources/cited_papers/index.json. Can force re-downloads and limit the number of new downloads.

• compile_latex

Runs LaTeX compilation (pdflatex/xelatex via latexmk) on the main document and returns success plus a log snippet. Supports specifying the entry .tex file and number of passes.

• read_pdf

Extracts text and metadata from a PDF filepath using pypdf, with page and character limits. Stores a JSON artifact for downstream use.

• read_pdf_from_citation

Resolves a citation key via resources/cited_papers/index.json to its PDF and extracts text (auto-downloads if needed). Returns extracted text, metadata, and the citation key.

---

TypeScript Bridge Tools

• summarize_text

Produces a concise natural language summary of provided LaTeX or text content. Accepts a maximum sentence count for the summary.

• suggest_bib_key

Suggests a stable BibTeX key based on authors, year, and title metadata. Useful for generating consistent citation keys.

---

Install (with uv)

From the repository root:

cd latex-mcp-server
uv tool install -e .

or `` uv tool install -e ./latex-mcp-server ``

This will make the console script latex-mcp-server available on your PATH.

If you prefer a one-off run without installing as a tool: ``bash uv run latex-mcp-server ``

MCP Configuration Snippet

Add to your mcp.json (VS Code / Claude Desktop user settings):

"latex-mcp-server": {
  "command": "uv",
  "args": ["tool", "run", "latex-mcp-server", "--workspace", "ABSOLUTE_PATH_TO_YOUR_PROJECT_ROOT"]
}

If you omit --workspace, the server will infer the current working directory.

Windows example

On Windows (PowerShell / Bash), if your paper_1 root is at C:\Users\User\projects\paper_1, use:

"latex-mcp-server": {
  "command": "uv",
  "args": [
    "tool", "run", "latex-mcp-server",
    "--workspace", "C:/Users/User/projects/paper_1"
  ]
}

Forward slashes are fine; they avoid the need to escape backslashes. Be sure to replace the placeholder ABSOLUTE_PATH_TO_YOUR_PAPER_1_ROOT — leaving a placeholder like WORKSPACE_FOLDER_PATH will cause the runtime error:

Workspace does not exist: C:\Users\User\WORKSPACE_FOLDER_PATH\PAPER_1

Alternative: Install once, then call directly

After running:

cd latex-mcp-server
uv tool install -e .

You can simplify the config to:

"latex-mcp-server": {
  "command": "latex-mcp-server",
  "args": ["--workspace", "C:/Users/User/projects/paper_1"]
}

Troubleshooting

| Symptom | Cause | Fix | |---------|-------|-----| | No solution found when resolving tool dependencies | The tool name latex-mcp-server isn’t published on PyPI and uv tried to resolve it as a dependency (often happens if you used a config field like requires or didn’t install locally) | Run uv tool install -e ./latex-mcp-server from repo root, then update your mcp.json to invoke the installed script directly | | Workspace does not exist: C:\\Users\\...WORKSPACE_FOLDER_PATH... | Placeholder path left unchanged | Replace with the real absolute path | | TypeScript runtime unavailable in tool results | Node.js not installed (only needed for TS bridge tools) | Install Node.js 18+ or ignore if you don’t need TS tools |

If you change TypeScript sources, rebuild and re-vendor (npm run build) before restarting the server.

TypeScript Bridge

The TypeScript sources live in ts_functions/src. Their compiled JavaScript output is vendored into the Python package under latex_mcp_server/ts_dist so users do not need npm or tsc just to run the server. To modify or add TS functions:

cd latex-mcp-server/ts_functions
npm install
npm run build

Then copy (or allow the provided helper script to sync) the dist/ contents into latex_mcp_server/ts_dist/.

Adding New Python Tools

Add functions in latex_mcp_server/functions/latex_ops.py (or new modules) and register them in server.py inside register_python_tools.

Citation-Aware PDF Reading

When a user asks about a paper (e.g. “Explain how \cite{smith2023model} differs from …”), the agent should:

1. Parse citation key(s) from the request.
2. Call read_pdf_from_citation for each key (fetches / downloads PDF and extracts text snippet + metadata).
3. Ground its answer on the extracted text (and optionally call summarize_text for a concise summary).
4. Indicate if extraction was truncated (check the truncated flag) and avoid over-claiming beyond available pages/characters.

The read_pdf_from_citation tool returns the same schema as read_pdf plus citation_key so multi-citation responses can attribute snippets correctly.

Auto-Updating index.json (Chosen Method: VS Code Tasks)

This project is configured to refresh resources/cited_papers/index.json before each LaTeX build using VS Code tasks (method #3 from the suggested approaches).

Added file: .vscode/tasks.json with three tasks:

• update-index: Runs a short Python one-liner invoking extract_bibliography on src/references.bib.
• latex-build: Invokes latexmk -pdf on main.tex.
• build+index: Sequentially runs update-index then latex-build.

How to Use

1. In VS Code, open the command palette and run “Tasks: Run Task”.
2. Choose build+index (recommend marking it as the default build task when prompted).
3. The BibTeX index will regenerate first; then LaTeX compilation proceeds.

If LaTeX Workshop cannot import the package (ModuleNotFoundError), a helper script latex-mcp-server/update_index.py is provided. The workspace settings have been updated to call: `` python latex-mcp-server/update_index.py `` so that a local editable install is not strictly required.

Make It the Default Build

When prompted by VS Code after first run, accept setting build+index as the default. You can also add to your workspace settings: ``jsonc // .vscode/settings.json (optional) { "latex-workshop.latex.recipe.default": "build+index" } ``

Why This Approach

• No background watcher process.
• Deterministic: index refresh tied explicitly to build command.
• Minimal overhead (BibTeX parsing is fast relative to PDF compilation).

If you later need an alternative (e.g., Makefile or on-demand freshness check), you can still layer it on; the tasks approach is non-invasive.

Adding New TS Tools

1. Create a new file in ts_functions/src/functions/ exporting a function.
2. Add it to the export map in ts_functions/src/index.ts.
3. Rebuild and sync dist output.
4. Register a Python wrapper in server.py using bridge.register_ts_tool.

Protocol Notes

This server follows the MCP JSON-RPC message patterns. If the mcp Python package updates its APIs, adjust imports accordingly.

License

MIT