ColTex

Quick start

You need a Linux or macOS machine with Docker (on Linux the installer can set it up), at least 2 CPU cores and 4 GB of memory. The images are prebuilt for x86-64 and ARM64.

curl -fsSL https://coltexeditor.github.io/install.sh | bash

Then open http://<server>:3690 and create the administrator account.

Why ColTex?

ColTex is an alternative to Overleaf for a research group, a lab or just yourself, running on a computer you control: a desktop in the lab, a home server or a cloud machine.

• Everything stays on your server. Projects, history, accounts and compile jobs never leave it. TeX runs in a sandboxed container without internet access.
• Built for writing together. Several people edit the same file at once, with comments, track changes, chat and version history.
• An AI agent that works on your project. It reads the files and the compile log, drafts changes, test-compiles them, and shows you a diff. Nothing is written until you apply it.
• One command to install, no ports to open. Tailscale and Cloudflare Tunnel are built in for remote access.

Features

Writing and compiling

• Source editor (CodeMirror) and a visual editor with Overleaf's table editor and figure dialog, toolbar, symbol palette, outline and breadcrumbs.
• LaTeX completions, a code check, equation preview, spell checking in many languages, word count, project-wide search and replace, and Vim or Emacs keys.
• Compiles with latexmk on a complete TeX Live. The engine (pdfLaTeX, XeLaTeX or LuaLaTeX) is picked from the document, Chinese included (ctex, xeCJK).
• PDF preview with SyncTeX in both directions, also in a separate window. Compile errors are underlined in the source, with plain-language hints.
• References: cite from your .bib files, search Crossref and arXiv (or paste a DOI or arXiv id) to add entries, or import a Zotero library or collection and keep it in sync.

Collaboration

• Real-time editing with everyone's cursors, comments with @mentions, track changes (accept, reject, and undo a decision), project chat and in-app notifications.
• Roles per project: owner, editor, reviewer (suggestions only) and viewer.
• Version history like git: automatic versions, versions saved with a message, and undoing one version's changes.
• Git access: clone a project, work on it locally and push back (personal tokens; edits made in the browser meanwhile are merged).

AI agent

• Ask about LaTeX, your project or your writing, or let it fix compile errors. It drafts changes in a private copy, test-compiles them and shows a diff; you apply all or part of it.
• Applying saves a version first and recompiles; if the build gets worse, the change is rolled back automatically.
• Providers: API keys (Anthropic, OpenAI, DeepSeek, Google, OpenRouter and more), subscription sign-in (Claude Pro/Max, ChatGPT, GitHub Copilot and more), or your own OpenAI-compatible server (Ollama, vLLM, LM Studio).
• Automatic model choice (a light or a strong model per message), access per group of people, usage per person, and usage limits per day, week or month.
• Saved chats per project, the context window shown as you go, and /compact to summarise a long chat.
• Skills: your own instructions (Agent Skills format) that the agent follows in all your projects, used with /name; write them yourself, import them, or create them with the AI in Settings.
• Permissions like Claude Code's: the AI agent's changes wait for your review (Manual), apply when it finishes (Auto), or it may also delete and move files (Full access).
• Your own agent (Claude Code, Codex, Cursor…) over MCP: one address for all your projects, with the same permissions (in Manual its edits become your suggestions).

Projects and administration

• Templates for IEEE, ACM, Springer LNCS, Elsevier, REVTeX and arXiv, and conversion of a paper from one journal template to another.
• Overleaf zip import and export, files linked from a URL or another project, tags, archive and trash.
• No public sign-up: accounts come from invitation links. One system administrator, and admins who each manage the team they invited.
• Monitoring, WebDAV backups (optionally encrypted), local or S3-compatible storage, light and dark themes, and an interface in 14 languages (English, Chinese, Japanese, Korean, Spanish, French, German, Portuguese, Italian, Russian, Turkish, Vietnamese and Indonesian).
• Account security: two-step verification with an authenticator app, a record of sign-ins with each address's place (looked up on the server), and protection against password guessing that locks an account for unfamiliar addresses, all without third-party services.

Install

If the download is cut off (curl: (35) Recv failure: Connection reset by peer, common when GitHub is hard to reach), run the command again.

It asks for the language and the TeX Live edition (arrow keys), gets the latest release into ~/coltex, downloads the prebuilt images with a progress bar, starts ColTex, and prints the addresses to open. The full output goes to ~/coltex/install.log. Then open http://<server>:3690 and create the administrator account. Remote access through Tailscale or Cloudflare Tunnel is turned on afterwards in Settings → Network.

The first start downloads TeX Live. Choose an edition:

Edition	Download	Disk (all of ColTex)	Free space to have
standard (default)	about 3 GB	about 12 GB	20 GB
full (--edition full)	about 8 GB	about 23 GB	35 GB

Both have every TeX Live package and compile the same documents, and the AI agent works with either. The standard edition leaves out the packages' manuals; the full edition adds the manuals and sources, which the AI agent can also read.

--edition and --lang answer the questions, or replace them where there is no terminal. --domain, --tailscale and the Cloudflare options set up remote access during the install, for scripted installs. Add options after bash -s --:

Option
--domain tex.example.org	HTTPS for a server with its own domain, through the bundled Caddy. Point the domain at the server and open ports 80 and 443
--tailscale	Put ColTex on your tailnet: prints a sign-in link and a QR code
--cloudflare-token TOKEN	Connect a Cloudflare tunnel you created for your own domain
--cloudflare-quick	A temporary public trycloudflare.com address, for trying it out
--edition full	The full TeX Live edition
--lang LANG	The language of the installer, coltex and the setup page: en, zh-CN, zh-TW, ja, ko, es, fr, de, pt-BR, it, ru, tr, vi or id (asked when not given)
--port 3690	Another port
--version v0.0.1	A specific release (main for the development branch)
--build	Build the images from the sources on this machine instead of downloading them
--images REGISTRY	Where to download the images from: aliyun (the same images on Alibaba Cloud), ghcr (ghcr.io/coltexeditor) or another mirror of ghcr.io. Without it, the installer uses whichever of the two answers faster from here (Docker's registry-mirrors only applies to Docker Hub, not to ghcr.io)

On a server reachable from the internet, don't sign in over plain http://: use --domain, Tailscale or a Cloudflare tunnel.

Manage and update

coltex status          # running? at which address
coltex logs            # the server's log (coltex logs compiler for the compiler)
coltex update          # to the latest release; your settings and data stay
coltex version
coltex edition full    # switch the TeX Live edition
coltex tailscale up    # or: coltex cloudflare quick | token
coltex stop | start
coltex reset           # delete all data and start again from the setup page (asks first)
coltex uninstall       # remove ColTex from this machine: containers, data, images, this folder (asks first)

Configure

Everything is set in the browser, under Settings in the user menu:

1. Manage users and Invitations: invite people with links (each works once, optionally for one email).
2. AI services: connect a provider. Then, in AI usage & access, put people in a group that may use it. Without a group nobody but the system administrator gets the AI agent.
3. Network: the addresses ColTex is reached at, and Tailscale and Cloudflare Tunnel.
4. Storage & backups: daily or weekly backups to a WebDAV folder (Nextcloud, a NAS…).

The help documentation is built in: Help documentation in the user menu. Server operations (environment variables, S3, restoring a backup) are in docs/operations.md.

Develop

Node 24 or newer, pnpm 11 and Docker.

pnpm install
cp .env.example .env
pnpm infra:up        # Postgres, Valkey and the TeX Live compiler in Docker
pnpm dev             # API on :3000, web app on http://localhost:5173
pnpm -r typecheck
pnpm -r test         # the server tests need Postgres from infra:up

Path
apps/web	The web app: React, CodeMirror 6, PDF.js
apps/server	The API: Fastify, accounts, projects, real-time collaboration, the AI agent
services/compiler	The compile worker in a TeX Live container
packages/*	Shared schemas, the log parser, LaTeX language support, review data and template conversion

The design is described in docs/architecture.md.

Acknowledgements

• Overleaf (AGPL-3.0): ColTex ports its log parser and hints, LaTeX language support, visual editor, table editor, spell checking and word count. Every ported file is listed in NOTICE.
• Pi (pi-ai, pi-agent-core): the model providers and the agent loop behind the AI agent.
• PilotDeck (AGPL-3.0): the provider health tracking and the automatic model choice are adapted from it.
• Yjs and Hocuspocus, CodeMirror, PDF.js, better-auth, shadcn/ui, TeX Live.
• Lobe Icons (MIT): provider logos. Trademarks belong to their owners and only identify the service.

License

AGPL-3.0. If you run a modified ColTex for other people over a network, you must offer them its source code. Third-party code keeps its own license; see NOTICE.