tex-mcp-web

Review a LaTeX paper from its rendered PDF while Claude Code or Codex edits the source.

If it helps your writing, a star is very welcome.

You read the PDF in a browser and comment on selected PDF text, selected source text, a section, or the whole paper. The agent reads those comments over MCP, edits the LaTeX, compiles once after the batch, and replies in the same thread.

Use the topbar's PDF, Source, and Split tabs to review the rendered paper, make a small source edit, or see both together. Drag the handle between the panes to adjust their sizes. Select source text and press + Comment to highlight exactly those characters, including text that wraps onto another display line. Source files follow the configured watch and ignore rules. Saving is explicit, and the editor refuses to overwrite a file changed by an agent or another editor after it was opened.

you:    select text -> write a comment -> press Call agent
                  |
agent:  read comments -> edit LaTeX -> compile -> reply
                  |
you:    read the rebuilt PDF -> comment again

Install

Python 3.10 or newer and a compiler on PATH (latexmk for LaTeX by default; pdflatex, xelatex, lualatex, and pandoc are also supported).

pip install "tex-mcp-web[mcp] @ git+https://github.com/MiiKiyoshi/tex-mcp-web"

Related MCP server: overleaf-mcp

Connect Claude Code or Codex

Register the MCP server once for the agent you use.

Claude Code:

claude mcp add --scope user tex-mcp -- tex-mcp

Codex:

codex mcp add tex-mcp -- tex-mcp

The agent starts tex-mcp itself; you do not run it separately. It also serves the review page for the project it starts in, so a paper needs no second process.

Set up a paper

Each paper needs a .tex-mcp-web.yaml at its root. Create it once in the paper directory:

cd my-paper
tex-mcp-web init --main main.tex

--main is the top-level source file that produces the PDF. init will not overwrite an existing file. It writes:

main: main.tex
watch:
- '*.tex'
- '*.bib'
- '*.md'
- '*.txt'
ignore:
- '*_backup.tex'
compiler: auto
auto_compile: false
port: 8765

To keep the paper out of the folder you start the agent in, put the config there and point dir at the paper:

cd my-project
tex-mcp-web init --main main.tex
tex-mcp-web config dir ../my-paper

The source, the PDF, and the comment store then stay under ../my-paper; the project folder holds only the config file.

Start Claude Code or Codex from the paper directory (or a subdirectory), then tell the agent:

Open the review page and listen for Call agent.

The agent opens the review page at the configured port and starts listening.

Use it

Drag over PDF text and write a comment; add a suggested wording in the replacement box when you have one. Use + Note for a whole-paper comment and the Sections tab for a section comment. Press Call agent when the comments are ready.

The agent reads them, edits the source, compiles, and replies in each thread. Resolving is yours: pick the threads whose edit satisfies you and resolve them from the page, or reopen one the agent got wrong. If an edit misses, reply in the same thread. Reference sets a thread aside to read again: it stays out of the open and resolved lists, still takes replies, and goes back to either with Reopen or Resolve.

For a long answer, the agent can ask read_comments(save=true) for a Markdown draft of the selected threads under .tex-mcp-web/drafts/, write its replies into the draft's Reply blocks, and send the file back with comment(action="reply", replies_file=...); the whole batch is applied together, and a draft made before a thread changed is refused. The draft also holds an Edit block for each of the agent's earlier entries; a changed block rewrites that entry in place (author and time kept, updated_at recorded). Inline, comment(action="edit", id=..., entry=..., text=..., updated=...) does the same by comment id, entry id and the thread's updated stamp, which read_comments reports.

If the agent restarts or stops receiving calls, ask it to listen for Call agent again. Calls made while it is disconnected stay queued.

Several papers at once

Give each paper its own port, then start an agent session in each directory:

cd paper-a && tex-mcp-web config port 8765
cd paper-b && tex-mcp-web config port 8766

They open at http://localhost:8765 and http://localhost:8766. A port already serving another paper is an error, not shared; after changing a port, restart that paper's agent session.

Configuration

tex-mcp-web config finds the nearest .tex-mcp-web.yaml from the current directory upward.

tex-mcp-web config                 # print the whole config
tex-mcp-web config port            # print one field
tex-mcp-web config port 8766       # change one field
tex-mcp-web config compiler xelatex
tex-mcp-web config watch 'main.tex,sections/**,*.bib'

Comma-separated patterns set watch and ignore.

More commands

tex-mcp-web compile          # compile once, without the review page
tex-mcp-web compile --json   # errors and warnings as JSON
tex-mcp-web goto Methods     # move a running viewer to a section, page (p2), or tex/intro.tex:47

Acknowledgements

tex-mcp-web is a hard fork of queelius/scholia at commit e6c7454 (version 0.6.1), independently developed since under the MIT license; see LICENSE. PDF viewing uses EmbedPDF and its PDFium WebAssembly engine, whose notices are in tex_mcp_web/static/embedpdf/LICENSE and tex_mcp_web/static/embedpdf/LICENSE.pdfium. Source highlighting uses Ace Editor builds 1.44.0 under the BSD-3-Clause license; its notice is in tex_mcp_web/static/ace/LICENSE.