Overleaf for AI agents (olmcp)

io.github.nipdepPublished by GitHub account nipdepv1.0.0-alpha.3Updated Nov 7, 2026

Overleaf for AI agents: review comments, tracked-change edits, new files and PDFs; no Git needed

VerifiedSTDIODesktop onlyProductivity & WorkflowDeveloper Tools

Overview

AI-generated overview

Local connector that lets an AI agent read Overleaf review comments, propose tracked-change edits, reply in threads and create text files.

What you can do
  • Read comments, threads, selected text and text files in an Overleaf project
  • Propose tracked-change edits in review mode, or apply a plain edit when a comment carries #direct
  • Reply in comment threads and revise a pending suggestion when asked for another version
  • Create new text files and missing folders, and compile the project to save the PDF locally
  • Skip or focus threads using #no-agent and #agent tags
When to use it
Useful when you review LaTeX papers or reports in Overleaf and want an agent to turn your margin comments into reviewable tracked changes without cloning the project or using Git. It suits iterative comment-and-revise loops where you stay the reviewer who accepts or rejects each edit.
Requirements
Node.js 20 or newer, an Overleaf account and an MCP-capable agent such as Claude Code or Codex; the GitHub-based install may also need git. Install the npm package olmcp and sign in with the overleaf_session2 cookie via olmcp auth, or set OVERLEAF_SESSION. OVERLEAF_BASE_URL is only needed for a self-hosted Overleaf instance. Desktop only.
Before you install
The Overleaf session cookie is a password-equivalent credential: store it via olmcp auth, never paste it into chats, issues or commits, and re-authenticate when it expires. Edits are attributed to your account, #direct changes apply immediately, and new files appear without tracked changes. An agent with shell access could read the stored cookie. The interface is unofficial and may break; the release is alpha, so start with a disposable project.

Installation

In SourceWeft

  1. Open Overleaf for AI agents (olmcp) in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

[Agent standing over a leaf]

Agent–Overleaf Connector (olmcp)

[tests]

olmcp lets an AI agent respond to review comments in an Overleaf project.

You comment in Overleaf. The agent reads the comment and the selected text, proposes an edit as an Overleaf tracked change, and replies in the same thread. You remain the reviewer: you accept or reject the edit and close the thread.

olmcp has no hosted backend: the connector runs on your computer and uses your Overleaf session. Your chosen AI client may still send MCP results to its model provider under that client's own privacy terms.

The workflow

text
1. You select text and leave a comment in Overleaf.2. The agent reads the comment and the live project.3. The agent proposes a tracked change and replies in the thread.4. You accept, reject, request another version, or resolve the thread.

By default, an edit is a proposal, not a silent replacement. Overleaf shows the original text as a tracked deletion and the proposed text as a tracked insertion.

Project status: 1.0.0-alpha.3. The core review flow has been used on real Overleaf projects; newer features such as pending-suggestion revision and file creation still need real-service testing. Start with a disposable project.

What the MCP can and cannot do

It canIt deliberately cannot
Read comments, threads, selected text, and text files in the projectAccept or reject a tracked change for you
Propose tracked edits in review modeResolve, reopen, close, or delete a comment thread
Apply a plain edit when you explicitly use #directOverwrite or delete an existing file
Reply in a comment threadEdit documents that use Overleaf's unsupported newer history format
Revise its own pending suggestion when you ask for another versionCreate images, PDFs, or other binary project files
Create a new text file and missing foldersWork continuously in the background or watch for new comments
Compile the project on Overleaf and save the PDF locallyGive the agent a separate Overleaf identity; changes use your account

The tool surface enforces these limits. However, an agent with unrestricted shell access could read the stored session cookie and contact Overleaf outside the MCP. See Security and trust.

Quick start

You need Node.js 20 or newer, an Overleaf account, and an MCP-capable agent such as Claude Code or Codex. You do not need Python, TeX, or Overleaf's paid Git integration. The current GitHub-based installation may also need the git command; normal use does not.

1. Install olmcp

bash
npm install -g github:nipdep/agent-overleaf-connectorolmcp --version

2. Sign in to Overleaf

Copy the overleaf_session2 cookie from your browser's developer tools, then run:

bash
olmcp auth --cookie "<cookie value>"

Treat this cookie like a password. Never paste it into an agent chat, issue, or commit. The setup guide shows where to find it and what to do when it expires.

3. Register the MCP server

Choose your agent, then restart it:

bash
# Claude Codeclaude mcp add --scope user olmcp -- "$(which olmcp)" mcp
# Codexcodex mcp add olmcp -- "$(which olmcp)" mcp

Windows PowerShell commands and other MCP clients are covered in Connect an AI agent.

4. Give the agent your project URL

No project folder or clone is required. After connecting the MCP, paste the Overleaf project URL into the request:

Address the comments in https://www.overleaf.com/project/&lt;project-id>.

If you prefer not to paste the URL each time, you can save the project ID in a working folder with olmcp init. See Choosing a project.

5. Address a comment

In Overleaf, select a sentence and comment “make this clearer.” Then tell your agent:

Address the comments in https://www.overleaf.com/project/&lt;project-id>.

Back in Overleaf, review the tracked edit. Accept or reject it, reply with another request, or resolve the thread yourself.

Everyday controls

When you want to…Do this in Overleaf
Propose a normal reviewable editLeave an ordinary comment; review is the default
Apply an edit without tracked changesAdd #direct to the comment
Keep the agent away from a threadAdd #no-agent
Focus the agent on particular threadsAdd #agent to those comments
Ask for another version of a pending suggestionReply in the same thread, for example “shorter please,” then run the agent again
Add a new section or bibliography fileAsk in a comment; the new file appears immediately, while the \input or bibliography edit is proposed separately
Finish a discussionResolve the thread yourself

Read Using olmcp for examples, modes, tags, file creation, and the complete MCP tool list.

Important safety details

  • Tracked changes are attributed to your Overleaf account because the connector uses your session.
  • A newly created file cannot be a tracked change. It appears immediately; the connector never overwrites or deletes files.
  • #direct changes are immediate. Use Overleaf History if you need to undo one.
  • The interface is unofficial and may break when Overleaf changes it.
  • The session cookie usually expires after a few days. Run olmcp auth again when olmcp doctor reports an expired session.

Documentation

Release changes are in the changelog. The retired Python workflow—Git sync, source markers, redline.tex, and GitHub Actions—remains on the python-legacy branch.

Logo files: PNG · SVG

License and credit

Apache 2.0. See LICENSE. The Overleaf protocol handling follows the approach used by olcli by Alexander Loth (MIT); see THIRD_PARTY_NOTICES.md.

Source: README.md at commit 0a07e78

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v1.0.0-alpha.3LatestNov 7, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts