overleaf-mcp
Enables AI assistants to read, edit, and compile LaTeX projects on Overleaf using session cookie authentication, bypassing the need for a paid plan.
README
overleaf-mcp
An MCP server for Overleaf that lets AI assistants (Claude, Cursor, etc.) read, edit, and compile your LaTeX projects directly — no more copy-pasting between the editor and your terminal.
No paid plan required. Overleaf's official Git integration is locked behind a paid subscription. This project instead authenticates the same way your browser does — with a session cookie — and drives the same HTTP and realtime endpoints the Overleaf web app itself uses. It works on the free plan.
⚠️ Before you use this
- This talks to undocumented, unofficial endpoints, including Overleaf's legacy Socket.IO realtime channel (used to resolve folder ids for uploads/deletes). It can break without notice if Overleaf changes its backend.
- Your session cookie is as powerful as your password for the duration of the session. Treat it like a secret: it's read from an environment variable and never leaves your machine, but don't paste it anywhere else.
- This is not affiliated with or endorsed by Overleaf/Writefull. Use it at your own risk and in line with Overleaf's Terms of Service.
Install
git clone https://github.com/NiccoloSalvini/overleaf-mcp
cd overleaf-mcp
uv sync
Get your session cookie
-
Log into overleaf.com in your browser.
-
Open devtools → Application (Chrome) or Storage (Firefox) → Cookies →
https://www.overleaf.com. -
Copy the value of
overleaf_session2(andGCLBif present). -
Build a
Cookie:header string:overleaf_session2=<value>; GCLB=<value>Easiest way in Chrome: open the Network tab, reload the dashboard, click any request to overleaf.com, and copy the full
Cookierequest header from Headers → Request Headers.
The cookie expires (Overleaf sessions typically last a few weeks). When tools start failing with an authentication error, repeat these steps and update the environment variable.
Configure
Set the cookie as an environment variable:
export OVERLEAF_COOKIE="overleaf_session2=...; GCLB=..."
Claude Code
claude mcp add overleaf --env OVERLEAF_COOKIE="overleaf_session2=...; GCLB=..." -- uv --directory /path/to/overleaf-mcp run overleaf-mcp
Claude Desktop (claude_desktop_config.json)
{
"mcpServers": {
"overleaf": {
"command": "uv",
"args": ["--directory", "/path/to/overleaf-mcp", "run", "overleaf-mcp"],
"env": {
"OVERLEAF_COOKIE": "overleaf_session2=...; GCLB=..."
}
}
}
}
Tools
| Tool | Description | Status |
|---|---|---|
overleaf_whoami |
Check whether OVERLEAF_COOKIE is set and still valid |
✅ verified live |
overleaf_list_projects |
List your projects (id, name, owner, last updated) | ✅ verified live |
overleaf_pull |
Download a project and extract it into a local directory | ✅ verified live |
overleaf_compile |
Trigger a compile, return status and error log | ✅ verified live |
overleaf_download_pdf |
Compile and download the resulting PDF | ✅ verified live |
overleaf_push_file |
Upload/overwrite a single file, creating remote folders as needed | ✅ verified live |
overleaf_push_dir |
Diff a local directory against the project and push changed files | ✅ verified live (calls push_file) |
overleaf_delete_file |
Delete a file from a project | ✅ verified live |
Every tool that takes a project argument accepts either the 24-character
project id or the exact project name.
All eight tools were live-tested 2026-07-19 against a real account, including
create/upload/verify/delete round trips on disposable test projects.
overleaf_list_projects/overleaf_pull needed a fix (Overleaf renamed a
meta tag). overleaf_push_file needed a bigger one: black-box guessing at
the upload request shape (field names, folder_id vs parent_folder_id,
with/without qqtotalfilesize) consistently got 422 invalid_filename, even
with a correct folder id read out of Overleaf's own React state. The actual
cause, found by reading Overleaf's own source
(services/web/app/src/Features/Uploads/ProjectUploadController.mjs in
overleaf/overleaf, AGPL): the
filename comes from a multipart form field name, not from a
qqfilename query string param the way the legacy Fine Uploader-era
endpoint mapping (and every third-party client that copied it, including the
original version of this one) assumed. folder_id and _csrf genuinely are
query params — only the filename field was wrong. Fixed in client.py.
overleaf_push_file, overleaf_push_dir, and overleaf_delete_file first
join Overleaf's realtime channel to resolve folder ids (the only way to get
them without the paid Git API) — expect these to take a few seconds longer,
and to be the first thing that breaks if Overleaf changes its backend.
How it works
- List / pull: the Overleaf dashboard (
GET /project) embeds the project list as JSON in a<meta name="ol-prefetchedProjectsBlob">tag, and a CSRF token in<meta name="ol-csrfToken">. Projects download as a zip fromGET /project/{id}/download/zip. Both are plain HTTP + HTML parsing. - Push / delete: uploading or deleting a file requires the internal
folder id, which Overleaf only exposes over its realtime channel (a
joinProjectSocket.IO event, using Overleaf's legacy pre-1.0 Socket.IO protocol). This project joins that channel, reads the returned folder tree, and then calls the same REST upload/delete endpoints the editor uses. - Compile:
POST /project/{id}/compilewith a CSRF token, same as the in-browser editor's build button.
Endpoint mapping originally traced from
moritzgloeckl/overleaf-sync
(MIT licensed), reimplemented here with httpx and typed models.
Development
uv sync --group dev
uv run pytest
uv run ruff check .
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。