Connecting ChatGPT and Claude to LaTeX: Overleaf MCP servers vs a hosted one
By Shihab Shahriar Antor · Updated August 2026 · 9 min read
Every Overleaf MCP server is unofficial. Each one bridges to Overleaf over its Git endpoint, and most run locally on your machine, so you need an Overleaf account and a terminal. letX runs a first-party hosted MCP server at api.letx.app/mcp: you connect it with OAuth, nothing installs, and it compiles on the server, so the assistant reads the real latexmk log.
The short version
There are two different things being called a LaTeX MCP server, and they are not competing for the same job. One group bridges an assistant into projects you already keep in Overleaf. The other gives the assistant an editor and a compiler of its own. Which one you want depends entirely on where your documents live.
| Overleaf MCP bridges | letX MCP | |
|---|---|---|
| Who publishes it | Independent developers | letX, first-party |
| Where it runs | Mostly on your machine (stdio) | Hosted, remote |
| What you install | Node, the server, a config file | Nothing |
| How it authenticates | An Overleaf session or Git token you paste | OAuth 2.1, revocable |
| Needs an Overleaf account | Yes | No |
| How it edits | Git commits against the project | Direct file writes, with version history |
| Can it compile | Some can, via Overleaf | Yes, on letX TeX Live workers |
| Does the assistant see the compile log | Varies | Yes, the real latexmk log |
What the Overleaf bridges do well
If your thesis is already in Overleaf, shared with a supervisor who is not going to move, then a bridge is the honest answer and we are not going to pretend otherwise. Projects such as OverleafMCP and overleaf-claude-mcp let an assistant read the document structure and push section-level edits back as real Git commits, which means every change is reviewable and revertible with tools you already know.
The costs are the ones that come with any unofficial integration. They depend on an interface Overleaf did not publish for this purpose and can change. Git access is a paid Overleaf feature on most plans. And a token pasted into a config file on your laptop is a credential with no expiry and no revoke button, sitting in plain text.
What a first-party hosted server does differently
letX runs the server itself, at https://api.letx.app/mcp. You add it as a connector, sign in once, and approve the scopes on a consent screen that names the client asking. There is no package to install, nothing running on your machine, and the grant is revocable from your account rather than by deleting a line from a JSON file.
The part that changes how it feels to use is compiling. The assistant does not guess whether the LaTeX is valid; it runs the build and reads what came back. When a document fails, it gets the actual latexmk log, with the file and line the error came from, and fixes that line. That loop is the difference between an assistant that writes plausible LaTeX and one that produces a PDF.
The twelve tools
- Read: search_templates, get_template, list_projects, list_project_files, read_project_file
- Write: create_project, create_project_from_template, update_project_file, compile_latex, compile_project
- Deep research: search and fetch, the pair ChatGPT and Claude research modes expect, so a research answer can cite a real template page rather than paraphrase a listing
Every write tool saves a version snapshot before it changes anything, so an edit an assistant made can be undone from letX history like any other. Nothing an assistant does is unrecoverable.
Setting it up
In Claude or ChatGPT, add a custom connector and give it the server URL. The consent screen tells you which client is asking, where it will be redirected, and which scopes it wants, and you approve or refuse per scope.
https://api.letx.app/mcpStep-by-step instructions per client are on Claude, ChatGPT and Claude Code.
Which one should you use
- Your documents are in Overleaf and staying there. Use a bridge. It is the only thing that reaches them.
- You want an assistant that can start a paper, not just edit one. Use the hosted server: it can open a project from any of 1,000+ templates and compile it.
- You do not want to run anything locally. Hosted, with OAuth.
- You want the assistant to fix its own compile errors. It needs to see the log, which means it needs to run the build.
Connect letX to the assistant you already use. Free on every account, including the free tier.
Set up the MCP serverFrequently asked questions
Is there an official Overleaf MCP server?
No. As of August 2026 Overleaf does not publish one. Every Overleaf MCP server is a third-party project that bridges to Overleaf over its Git endpoint, so it depends on an interface Overleaf did not design for this and can change without notice.
Can ChatGPT edit my LaTeX and compile it?
Yes, if it is connected to a server that can do both. letX exposes compile_latex and compile_project over MCP, so the assistant runs the real build on letX workers and gets the PDF back, or the latexmk log when it fails. Most Overleaf bridges can read and write files; whether they can compile varies by project.
Do I need a paid plan to use the letX MCP server?
No. The MCP server is free on every letX account, including the free tier. What it can do is bounded by your account, so the same limits apply as when you use letX directly.
Is it safe to let an AI assistant edit my documents?
Every write tool takes a version snapshot before it changes anything, so an assistant edit can be reverted from letX history like any other. Access is granted per scope on a consent screen and can be revoked from your account at any time, which is not true of a Git token pasted into a config file.
Does letX train AI models on my documents?
No. We do not use your private documents to train AI models without your explicit, opt-in consent, and there is no such programme today. When you use an AI feature the relevant text goes to a model provider only to generate your result.
Why does the assistant need a compiler at all?
Because LaTeX errors are the hard part, and they only exist at build time. An assistant that cannot compile is guessing at whether its output works. One that can compile reads the failing line out of the latexmk log and fixes that line, which is a different quality of help.
Related
Written by Shihab Shahriar Antor , AI Engineer & Founder of Shahriar Labs. Builder of LetX (collaborative LaTeX) and QuantumSketch (AI STEM video).