refereekit

com.inferweavePublished by inferweave.comv1.0.2Updated Oct 31, 2026

Confidential peer review that checks a paper's citations. Set up from the install page first.

Overview

AI-generated overview

Local peer-review assistant that checks a draft's quotations, pages, equations and figures against a submitted PDF and helps draft a referee report and editor…

What you can do
  • Start or resume a review session on a named PDF, returning a session reference and page count
  • Check a draft against the paper with deterministic checks and get verdict counts and reason codes
  • Read session status as codes and counts, per session or across all sessions
  • Open private views on the referee's own screen for the review, drafts, flags, notes, finishing checks and tutorial
  • Run the final finishing checks on a session
  • Delete a session or remove/reinstall the agent integration, each confirmed on the referee's screen
When to use it
For referees reviewing a manuscript who want an assistant to verify that every quotation, page, equation and figure cited in their answers or drafts actually appears in the submitted PDF, and to help draft a report and editor letter from claims that checked out.
Requirements
Runs as a local process over stdio, installed from PyPI with uv (uv tool install refereekit) and started with refx chat paper.pdf; needs a terminal and uv. A model backend is configured in the private view; anthropic, bedrock and vertex need one extra install line. A practice tutorial needs no key.
Before you install
The manuscript is confidential: keep its text, quotations, verdicts, notes and credentials out of tool arguments and chat, and pass only file paths and session references. The paper is never given to the agent's model, but anything typed in the agent's chat does reach it, so avoid the paper's title. A clean finish is not approval; the verdict stays with the referee. Deleting a session or removing the integration requires confirmation on the referee's screen.

Installation

In SourceWeft

  1. Open refereekit 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

refereekit

refereekit helps you referee a paper with a language model. It checks every quotation, page, equation and figure that an answer or a draft cites against the submitted PDF. It then drafts a referee report and an editor letter in your style, from claims that checked out. You edit them. The verdict is yours.

You need a terminal and uv. Two commands start a review:

text
uv tool install refereekitrefx chat paper.pdf

In the chat, /model access sets your model. For anthropic, bedrock and vertex it names one more install line. /tutorial runs a practice review that needs no key. Your terminal is the private view.

You can also review in pi, Claude Code, Codex or Claude Desktop, or in the ChatGPT desktop app as a preview. Paste the setup prompt into your agent, which sets refereekit up. refereekit never gives the paper to the agent's model. What you type in the agent's chat does reach it, so keep the paper's title out. You read the paper, and type your judgment, in a private view.

Confidentiality first. A manuscript under review goes only to a model backend, after you attest what its account keeps. It never enters this repository. See Confidentiality.

Documentation: start with Before you start, then pick a reading path from the documentation index.

Citation

If you use refereekit in your research, please cite the following paper:

bibtex
@inproceedings{yen2026models,    title={Models Quote, Code Verifies, Humans Judge: Lessons for AI-Assisted Peer Review},    author={Tzu-Chi Yen},    booktitle={NeurIPS 2026 Workshop on AI-Native Academia},    year={2026}}

License

MIT. See LICENSE.

Source: README.md at commit 6dd6b3e

Tools

8

refx_start

refereekit: start

Start a refereekit session on a PDF the referee named, or resume one. Returns the session reference and page count, never the paper's text.

Parameterspdfrequiredvenuesession

refx_status

refereekit: status

Where one refereekit session stands, or a line for each session. Codes and counts only. PASS: the check's own guarantee holds, and nothing more. Quotation: the quoted words occur on the named page. Figure: a caption carries its label; what it shows is unchecked. FLAG: the page exists; its wording or an equation's label is unchecked. FAIL: the words are not on that page, the page does not exist or was not named, or a figure label was not found. It never shows the paper lacks them. A draft citing it is flagged. UNANCHORED: a quotation or equation with no clear page; unchecked. A clean finish: nothing is missing and nothing FAILs. It is not approval.

Parameterssession

refx_check

refereekit: check

Check one of the session's drafts against the paper, without a model. Returns verdict counts and reason codes; the referee reads the details in the window. PASS: the check's own guarantee holds, and nothing more. Quotation: the quoted words occur on the named page. Figure: a caption carries its label; what it shows is unchecked. FLAG: the page exists; its wording or an equation's label is unchecked. FAIL: the words are not on that page, the page does not exist or was not named, or a figure label was not found. It never shows the paper lacks them. A draft citing it is flagged. UNANCHORED: a quotation or equation with no clear page; unchecked. A clean finish: nothing is missing and nothing FAILs. It is not approval.

Parameterssessionrequiredartifactrequired

refx_finish

refereekit: finish

Run the last deterministic checks on a refereekit session. Nothing is submitted or approved. PASS: the check's own guarantee holds, and nothing more. Quotation: the quoted words occur on the named page. Figure: a caption carries its label; what it shows is unchecked. FLAG: the page exists; its wording or an equation's label is unchecked. FAIL: the words are not on that page, the page does not exist or was not named, or a figure label was not found. It never shows the paper lacks them. A draft citing it is flagged. UNANCHORED: a quotation or equation with no clear page; unchecked. A clean finish: nothing is missing and nothing FAILs. It is not approval.

Parameterssessionrequired

refx_draft

refereekit: draft

Open the review window at the report or the editor letter. The referee types the section lengths or the answers for the editor there and reads the draft there; refx_status shows when it is written.

Parameterssessionartifactrequired

refx_open

refereekit: open a view

Open one of the referee's private views in a window on the referee's own screen: review (the questions, the verdict and both drafts), drafts (read, check and edit a draft in the referee's editor), flags (what needs attention), note (venue notes), finish (the finishing checks), tutorial (a practice review) or setup (the reviews folder's page: model access, where the referee enters the backend, its key and the attestation, and the OpenReview account; it needs no session). The referee types and reads everything there; the result says whether the window opened.

Parametersviewrequiredsession

refx_close

refereekit: delete a session

Delete one refereekit session. The referee is asked on their own screen, with what the deletion destroys; the session goes only on their yes there, never on a tool argument. If the session holds venue notes, pass notes_to (a database path the referee named) or drop_notes as the referee decided. The result is the dry run's counts, the deletion's, or the window that asks.

Parameterssessionrequirednotes_todrop_notes

refx_setup

refereekit: remove or reinstall the agent integration

Remove refereekit's agent integration from this reviews folder (uninit), or reinstall it here (reinit). The referee is asked on their own screen, with what it would change; it happens only on their yes there, never on a tool argument. Sessions and the referee's own files are never touched. The result is the dry run's counts, the change's, or the window that asks.

Parameterstoolsactionrequiredprofile

Version history

1
  1. v1.0.2LatestOct 31, 2026

Score

90% 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
  • 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 listedRequired15 pts
  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts