Compile LaTeX From Inside ChatGPT and Claude
An assistant writing LaTeX in a chat window has no idea whether it compiles. That is the whole problem. It produces something well-formed, you paste it into an editor, and find out that the table needs a package nobody loaded. Connecting a real compiler closes the loop: the assistant writes the file, builds it, reads the log, and fixes what broke before it answers you.
The mechanism is the Model Context Protocol, and the setup is one URL.
Connect it, in whichever client you use
- 1Open Settings, then Connectors.
- 2Choose to add a custom connector.
- 3Paste the server URL: https://api.letx.app/mcp
- 4Sign in with your LetX account when the consent screen appears.
- 5Approve the tools you want it to use.
Custom connectors are not available on every ChatGPT plan. If the option is missing, the plan is why.
The server is remote and speaks OAuth, so there is nothing to install and nothing to keep running. The same URL works in every client above.
What changes once it is connected
The visible difference is that you stop copying things. Ask for a conference poster and you get a project in your account with a compiled PDF, not a code block to paste somewhere. Ask for a fix and the assistant can check its own work, which is the part that matters: a compile either succeeded or it did not, and that answer does not depend on the model being confident.
The less visible difference is that the catalogue comes with it. The assistant can search over a thousand templates and start from a real one, so "make me a thesis in the Cambridge format" begins from the Cambridge layout rather than from an article class and a guess about margins.
What it will not do
It cannot reach outside your projects, and it has no shell. The tools it is given are deliberately narrow: search templates, create a project, read and write files inside it, compile. That is enough for the work and small enough that a mistake is recoverable, which is the trade every tool surface exposed to a model has to make. Nothing in it deletes a project.
If you would rather stay in the editor
MCP is for people who already live in an assistant. If you would rather have the assistant come to the document instead, that is Lexi, which sits in the editor itself and proposes diffs you accept or reject. The prompts that actually change the document applies to both, because the thing that decides whether you get an edit or an essay is the sentence, not the surface.
Frequently asked questions
What is an MCP server, in practical terms?
A way for an AI assistant to use a tool that lives somewhere else. Instead of the model writing LaTeX into the chat and hoping, it calls a real service that creates the project, writes the files, and runs a compiler, then reports what happened.
Do I need to install anything to use LetX from ChatGPT or Claude?
No. It is a remote server reached over HTTPS with OAuth, so you paste one URL into the connector settings and sign in. There is nothing running on your machine and nothing to keep updated.
What can the assistant actually do once connected?
Search the template catalogue, create a project from a template, read and write project files, compile to PDF, and hand back a link to the result. It works on your real projects, so what it produces is in your account afterwards.
Is my document sent to the model provider?
The parts the assistant reads pass through the assistant, the same as anything you paste into a chat. What changes is that the file itself lives in your LetX project rather than in the conversation, and letx.app declares ai-train=no, so your documents are not training data for us.
Ready to try LetX?
Join researchers and students focusing on their work, not the tools.
Get Started Free
Written by Shihab Shahriar Antor , AI Engineer & Founder of Shahriar Labs. Builder of LetX (collaborative LaTeX) and QuantumSketch (AI STEM video).