WebLatexMCP

Edit, compile, and sync your Overleaf projects with Claude.

       

Works with  

Pending      

---

WebLatexMCP lets Claude work on your paper the way a co-author would — rewriting the paragraph you point at, compiling to find out what actually broke, checking a citation in databases like Crossref or OpenAlex, looking at the typeset page before it tells you the figure fits. You review every change, and nothing reaches Overleaf until you say so.

Highlights

• 🗂️ Any project, with or without a remote — Overleaf, GitHub, or a local folder.
• 🧪 Local compiles — two backends run on your machine: latexmk, which is what Overleaf runs, or tectonic.
• 👀 Live PDF viewer + review comments — a viewer that hot-reloads on every compile, in a browser or a VS Code tab. Select text in the PDF to leave a note, and Claude applies it at the right source line.
• 📚 Citations checked, not trusted — API (CrossRef, OpenAlex) calls to verify, or add citations.
• 🧑‍⚖️ Pre-submission peer review — a multi-model reviewer panel, merged into one verified review (how it works).
• 🧩 Bundled Claude Code skills — project cleanup, typo hunting, writing-guide review, citation audits, bibliography normalization.
• 🔐 Tokens stay in memory — never written to .git/config, and scrubbed from all output.

Install

Claude Code (CLI or the VS Code extension)

Point Claude at this repo — the plugin registers the server and the skills in every session, from any directory:

👽 Please setup the following MCP server https://github.com/elias-ramzi/WebLatexMCP

Claude Desktop — one-click extension

Download web-latex-mcp.mcpb from the latest release and drag it onto the Claude Desktop window (or Settings → Extensions → Install Extension). No cloning, building, or JSON editing — Desktop shows a short, all-optional form (tokens, clone folder).

Then, from the chat

Hand over a token and add your project. The token goes through a local page straight into your OS keychain — never through the conversation:

👽 Open the credential portal for my Overleaf token, then add https://git.overleaf.com/… as "thesis".

Everything else — the npm package on its own, per-OS setup, TeX backends, other clients, iPad and browser — is in the installation guides.

Skills

Task-specific skills that drive the tools — each stops at the diff, so nothing is committed or pushed unless you ask:

• /format-latex-project — split the main file into per-section \inputs, move each figure/table into its own \input file, and reflow to one sentence per line.
• /arxiv-clean-project — run arxiv-latex-cleaner to strip comments and draft macros (\todo, notes) for arXiv, as a separate submission copy or applied in place.
• /verify-citations — audit a document's references against DBLP, Crossref or OpenAlex, flag discrepancies, and write a local audit report.
• /format-bibliography — deduplicate, normalize cite keys, harmonize venues, propagate renames into \cites.
• /proofread-document — hunt typos (spelling, doubled words, agreement, punctuation, LaTeX escapes).
• /review-writing-guide — review the paper against the writing guide and report prioritized suggestions with a concrete rewrite each.
• /peer-review — pre-submission review of an ML paper: summary, strengths, weaknesses, minor weaknesses, questions, typos — in Claude Code, /review-paper runs it as a Sonnet/Opus/Fable panel merged by a Fable triage.
• /summarize-paper — write/update a small local summary of the paper (git-excluded) so future sessions start fast.

Every bundled skill, and every command and agent the plugin ships, is scanned on each pull request by NVIDIA's SkillSpector (static rules, --no-llm), and any finding fails CI. Today every skill scores 0/100 with no findings.

How you get them depends on the client:

• Claude Code — install the plugin, and Claude picks a skill up when your request matches it. The plugin also brings the paper commands (/review-paper, /hunt-typo, /format-latex, /review-writing, /rewrite-mode, named /web-latex-mcp:<command>) and the agents they dispatch.
• Any MCP client — nothing to install: every skill ships with the server as an MCP prompt, in the client's prompt menu.
• Claude Desktop / claude.ai — for the same automatic behavior, zip each folder under .claude/skills/ and upload them under Customize → Skills.

See the skills guide for what each skill does, step-by-step installation, and the two ways a skill runs.

Documentation

Everything above in depth: Configuration, Tools, Skills, Concurrency, Writing guide

Contributing

This repo accepts pull requests — bug reports, feature ideas, docs fixes, and code changes are all welcome. See CONTRIBUTING.md for how to get set up, run the local gate, and open a PR.

Telling us how a session went is a contribution too: run the /session-feedback skill at the end of a session and it writes up what to improve as ready-to-file issue bodies (details).

A note on maturity: this project is largely vibe-coded, so treat it as best-effort rather than battle-tested. Robustness isn't guaranteed — expect rough edges, and please report them. It has been mostly tested on these setups: VS Code + Claude Code extension, the Claude Code CLI, and Claude Desktop for macOS.

License

MIT