Install

You set refereekit up by pasting one message, the setup prompt, into your agent. Part 1 needs no key and no account. Part 2, model access, can wait until your first real review.

What you need

A Mac with Apple silicon and macOS 14 or newer, or Ubuntu 24.04 on x86-64.
uv, installed as its page describes. It is the one thing you install yourself.
An agent: pi 0.99.1, Claude Code 2.1.295 or Codex 0.159.2, or newer. Claude Desktop and the ChatGPT desktop app work too. They run no hook guard, so they protect the paper less. Read Desktop apps: no hooks before you review a real paper in one.

Part 1: get it running

The setup prompt

Paste this message into your agent, as it is:

Please set up refereekit on this computer for me. refereekit lets me referee a
paper confidentially inside an AI agent like you. Work through the steps below
in order.

Rules for the whole setup:

- Before each install and each write, tell me the exact command and wait for my
  approval. Run each one under your own permission prompt.
- Never edit shell startup files, and never install Python or any other
  language runtime yourself; uv provides the Python refereekit runs on. Do
  not run a command that changes shell startup files, such as
  `uv tool update-shell`.
- Never ask me for a key, a password or any text of a manuscript, and never
  read a `.env` file. I set model access myself later, in a private view, not
  in this chat.
- If something is missing, stop and tell me what is missing. Do not work
  around it.
- If you cannot run commands on this computer, stop and tell me so. Setup then
  needs an agent that can, such as Claude Code or Codex on this computer, given
  this same message and the host name of the app I will review in.

1. Check the prerequisites. Run `uv --version`. Name the host you are
   running in and its version where it shows one. The hosts refereekit
   supports are pi, Claude Code, Codex, Claude Desktop and the ChatGPT
   desktop app. If uv or a supported host is missing, stop and tell me which.
   Do not check for Python: uv finds or provides one that suits refereekit.

2. Install refereekit with uv, with its model and OpenReview support. Until
   version 0.2.0 is on PyPI, install it from its Git repository:
   `uv tool install 'refereekit[llm,openreview] @ git+https://github.com/junipertcy/refereekit'`
   Once 0.2.0 is on PyPI, the command is
   `uv tool install 'refereekit[llm,openreview]'`. If I say I will use
   Amazon Bedrock or Google Vertex AI for the model, add
   `--with 'anthropic[bedrock]'` or `--with 'anthropic[vertex]'` to that
   command; the same command with that addition adds it later. Then run
   `refereekit --help`. If `refereekit` is not found, stop and tell me that
   uv's tool folder is not on the PATH.

3. Ask me where to keep my reviews, for example a new folder named `reviews`
   in my home folder. Then run `refereekit init <folder> --tools <host>`,
   where `<host>` names the app I will review in: `pi`, `claude` for Claude
   Code, `codex`, `claude-desktop` for Claude Desktop, or `chatgpt` for the
   ChatGPT desktop app. If it reports a conflict or an error code, stop and
   tell me the code.

4. In that folder, run `refereekit doctor` and report each check's name and
   status. The tutorial needs no model access, so a check for a real review
   that is not configured yet is expected.

5. Tell me what comes next:
   - Whether to restart or trust anything. pi, Claude Desktop and the ChatGPT
     desktop app load refereekit when they restart. Claude Code and Codex load
     it when started in the reviews folder: Claude Code asks me to approve the
     folder's server, and Codex asks me to trust the folder and to approve
     its hooks with `/hooks`.
   - How to set model access. I ask you to open it. In pi, that is the
     `/refx-model` command. In Claude Code, Codex, Claude Desktop and the
     ChatGPT desktop app, you call `refx_open` with view `setup`. I choose the
     backend and enter its credentials in that private view, never in the
     chat.
   - That I can ask you for the tutorial, an offline practice review, at
     any time.

The prompt needs an agent that can run commands. Claude Desktop and the ChatGPT desktop app cannot. For them, paste it into Claude Code or Codex on the same computer, and name the app you will review in.

What you approve

You approve each command in the agent’s permission prompt: the install, a reviews folder where you choose, the files that connect refereekit to your agent, and a readiness check. The agent asks for nothing secret.

The reviews folder holds your settings, and one private directory per review under work/. Edit its STYLE.md, the voice of your drafts, in any text editor.

What each host needs next

Host	Then
pi	start pi again in the reviews folder
Claude Code	start it in the reviews folder, and approve its `refereekit` server
Codex	start it in the reviews folder, trust the folder, and approve its hooks with `/hooks`
Claude Desktop or the ChatGPT desktop app	quit it and open it again

The practice run

Ask your agent for the tutorial: /refx-tutorial in pi or Claude Code, $refx-tutorial in Codex, or in plain words in a desktop app. To upgrade later, ask your agent to upgrade refereekit and update your reviews folder.

Part 2: model access and OpenReview

A real review sends the manuscript to a model backend. You choose it in Model access, a private view. In pi, type /refx-model. Elsewhere, ask the agent to open model access, and a page opens in your browser.

There you choose a deployment, the route that carries the manuscript, and enter its settings. Last, you attest that the account behind it keeps nothing:

Deployment	Carries the manuscript to	You attest
`anthropic`, the default	Anthropic’s API	Your organization has a zero-data-retention arrangement.
`bedrock`	your AWS account	Your AWS account has no model-invocation logging.
`vertex`	your Google Cloud project	Your project’s logging and retention settings permit it.
`claude-cli`	the Claude Code CLI on this computer	The Claude Code login that refereekit runs as `claude -p` sits on terms and settings that retain nothing.

vertex also needs a model id. Then ask the agent whether refereekit is ready for a real review. That check sends nothing.

The attestation

Without the attestation, nothing is sent. It is your statement, and refereekit cannot check it. For claude-cli, consumer and subscription plans generally offer no zero-retention terms. See the attestation.

Your .env

What you type in Model access goes to one file, the .env in your reviews folder, readable by you alone. It never reaches the chat. pi has no password field, so a key you type in pi shows on your screen.

OpenReview

Fetching a submission from OpenReview needs your user name and password. Enter both in the page’s “OpenReview account” section, or in /refx-model in pi. Reviewing on OpenReview has the rest.