Compile LaTeX projects to PDF, for AI agents and for people. Upload a .zip, .7z, .tar,
.tar.gz, .tar.bz2 or .tar.xz archive of a project (with its BibTeX/biblatex
bibliography) and get back the PDF and the compilation log.
- Full TeX Live (every package, as on Overleaf), compiled with
latexmkusing pdfLaTeX (default), XeLaTeX or LuaLaTeX; BibTeX or Biber run automatically. Like Overleaf, Ghostscript is there (EPS figures) and Java is not. - Main file detection: the only
.texfile at the top of the project, or the only one there with a\documentclass. Otherwise you name it. - Queue: jobs from all users run one at a time, strictly in submission order, while the web server stays responsive.
- Per-user tokens and storage, a job list with status and queue position, downloads of the archive, PDF and log, and deletion of one or all jobs.
- Sandboxed compilation: documents cannot read other users' files or the tokens, and cannot use the network.
- A web page to drop an archive, pick the main file and the engine, and follow the job list (point at the ? of a failed job to see its errors), with the server's settings and a light or dark theme.
- An HTTP API for programs, and an MCP server through which AI agents such as Claude Code, Codex and Antigravity compile projects, read the errors and download the PDFs by themselves.
Built with Python, FastAPI and uvicorn on the TeX Live image. SQLite keeps the job queue.
docker build -t texdrop . # about 6 GB: the full TeX Live
docker run -d --name texdrop \
--restart unless-stopped \
-p 8000:8000 \
-e TZ=Europe/Athens \
-e AUTO_DELETE_DAYS=15 \
-v texdrop-data:/data \
texdrop
docker exec texdrop python -m texdrop adduser alice # prints alice's tokenThen open http://your-host:8000 and log in with the token. --restart unless-stopped
starts TeXdrop again after a crash or a reboot. Extra hardening (--cap-drop ALL,
--security-opt no-new-privileges) and other options are in
Deployment. Add users the same way, at any time;
python -m texdrop reset-token alice gives alice a new token if theirs is lost or leaked.
To let AI agents use TeXdrop, add -e ENABLE_MCP=1 and connect them as
MCP.md shows. For the full sandbox, the host needs
Linux 6.12 or newer.
Mount the volume at
/data: that is where TeXdrop keeps the users, the jobs and the log. To keep them in a directory on the host instead, give it to UID/GID 61001, the user TeXdrop runs as, first:sudo chown 61001:61001 /your/dir. See Deployment.
It also works behind an HTTPS reverse proxy as it is. There, set
PUBLIC_BASE_URLto the address users reach (such ashttps://texdrop.example.com), so that browsers may send changes only from pages at that address: see Reverse proxies.
- Deployment: building, running with or without Docker, the data directory, reverse proxies, container limits and logs.
- Configuration: the users, every environment variable, and the fixed limits.
- API: every endpoint, how jobs behave, and examples in the shell and Python.
- MCP server: connecting AI agents, the tools and the transfer links.
- Security: the sandbox and the other protections, older kernels, and how to report a vulnerability.
- Development: how it works, the tests, and updating the dependencies.
