Environment variables

You set what a review needs in model access, never by hand in a shell (see Install). Model access saves it in your reviews folder’s .env. This page lists the names it writes, for when you want to read that file or edit it yourself.

Where they come from

refereekit reads each name from the environment its agent started with first, then from the reviews folder’s .env. The environment wins. It reads the file when each step starts, so a change counts from the next step, with no restart. It reads no .env outside your reviews folder.

The deployment

REFEREEKIT_BACKEND names the deployment, the service that carries the paper to the model backend. It is one of four. Any other name is refused rather than replaced, so a typo cannot send the paper somewhere you did not choose. Left empty, it means anthropic.

Deployment	What it needs	Default model
`anthropic`	`ANTHROPIC_API_KEY`	`claude-opus-4-8`
`bedrock`	`AWS_REGION` or `AWS_DEFAULT_REGION`, and `AWS_PROFILE` or AWS keys	`anthropic.claude-opus-5`
`vertex`	`CLOUD_ML_REGION`, `ANTHROPIC_VERTEX_PROJECT_ID`, and Google credentials	none: you must name one
`claude-cli`	the Claude Code program, installed and signed in	your login’s own default

anthropic is the default only because a referee with an API key already has it. It is not the preferred choice. The vertex row is written from the provider’s code, and has not been run against a real project.

Three more names apply to every deployment:

REFEREEKIT_MODEL names the model, in place of the default.
REFEREEKIT_EFFORT sets how hard the model thinks, from low to max. Only claude-cli reads it.
REFEREEKIT_ZERO_RETENTION=1 is your attestation that the account keeps nothing. Without it, refereekit sends nothing. See The attestation for what it asserts on each deployment.

OpenReview

OPENREVIEW_USERNAME and OPENREVIEW_PASSWORD are your OpenReview login. Only fetching from OpenReview reads them. An empty value counts as missing.

Your editor

“Open it in your editor” opens a draft in the editor that VISUAL names, else EDITOR. In the review window it must open a window of its own, such as code --wait. With neither set, the draft opens in the app your desktop uses for text files. In pi, /refx-drafts offers your editor beside its own.

These come from the environment your agent started with, not from .env. A desktop app starts refereekit without your shell’s settings, so set them where the app can see them, or use the default app.

Your files

Three files in your reviews folder can be named elsewhere, by a variable in the environment your agent started with:

REFEREEKIT_STYLE: your style guide, in place of the folder’s STYLE.md;
REFEREEKIT_QUESTIONS: the shared list of starter questions;
REFEREEKIT_VENUE_POLICY: your venue policy, in place of the folder’s venue-policy.toml.

Most referees need none of them.

Keys stay out of the chat

Never put a key or a password in the chat. No tool takes one, and a call that tries is refused (see Credentials).