Configuration

Rehearsal reads its settings from environment variables. A server also reads a file named .env in the folder it starts in, and a value in the real environment wins over the file. With Docker Compose, the settings are in infra/deploy.env: see Install with Docker Compose.

A person who only uses a server needs the first section. The rest is for the person who runs one.

Client

Setting	Default	What it changes
`REHEARSAL_URL`	`https://rehearsalkit.xyz`	The server that the CLI, the Python client and the MCP server talk to. It takes precedence over a saved sign-in.
`REHEARSAL_API_KEY`		The key they use. It takes precedence over a saved sign-in.
`REHEARSAL_CONFIG`	`~/.config/rehearsal/config.json`	Where `rehearsal login` saves the sign-in.
`REHEARSAL_DEBUG`		Set it to anything to make the CLI log each HTTP request.

Database and files

Setting	Default	What it changes
`REHEARSAL_DATABASE_URL`	a local PostgreSQL on port 5433	The PostgreSQL database, as `postgresql+psycopg://user:password@host:port/database`.
`REHEARSAL_DB_POOL_SIZE`	`10`	Connections that each process keeps open.
`REHEARSAL_DB_MAX_OVERFLOW`	`20`	Extra connections that each process may open under load. A managed database limits the total of all processes.
`REHEARSAL_DATA_DIR`	`./var`	The folder for application sources, world snapshots and exported files. The API and every worker must see the same folder.
`REHEARSAL_SOURCE_ROOTS`	`./examples`	Folders, separated by commas, that an application may name as a local source.
`REHEARSAL_DOWNLOADS_DIR`	`downloads` in the data folder	A folder that the server offers at `/downloads`, for the client package.
`REHEARSAL_ENV_FILE`	`.env`	A file of settings to read at start. A value in the real environment wins over the file.

Setting	Default	What it changes
`REHEARSAL_SESSION_SECRET`	a development value	Signs browser sessions. Set a long random value on any server that others can reach.
`REHEARSAL_SIGNUP`	`open`	Any visitor can create an account and an organisation. `closed` turns that off.
`GOOGLE_CLIENT_ID`		A Google OAuth web client id. With it, the console offers "Sign in with Google".
`REHEARSAL_ADMIN_EMAILS`		Operator admins, separated by commas. They see every workspace and can set any plan. Only a sign-in with Google counts.
`REHEARSAL_PUBLIC_URL`	`http://localhost:7800`	Where browsers reach the server. Payment pages return here.
`REHEARSAL_DOCS_URL`		The documentation site. `/docs` on the server redirects there; without it, `/docs` opens the API reference at `/api/docs`.

Models

Setting	Default	What it changes
`REHEARSAL_MODEL_COMPILER`	a chain of three hosted models	The model chain that builds worlds.
`REHEARSAL_MODEL_AGENT`	the same chain	The model chain for agents that have no model of their own.
`REHEARSAL_MODEL_PARTICIPANT`	a chain of two smaller models	The model chain that plays the simulated customer.
`REHEARSAL_MODEL_IMPROVER`	the compiler's chain	The model chain that proposes a better agent.
`REHEARSAL_PRICES`	`{}`	Prices for models whose provider reports none: JSON of `provider:model` to `[input, output]` in US dollars for a million tokens.
`REHEARSAL_PRICING_FILE`	`pricing.yaml`	A file of the same prices, each with its source and date.

Model chains

A chain is a list of provider:model, separated by commas. The server tries them in order, and moves to the next when a provider fails or is not configured.

REHEARSAL_MODEL_COMPILER=nebius:deepseek-ai/DeepSeek-V4.1-Flash,google:gemini-3.5-flash

Providers

A provider works when its key is set. Every provider is reached through an OpenAI-compatible API.

Provider	Key	Address
`nebius`	`NEBIUS_API_KEY`	`https://api.tokenfactory.nebius.com/v1`
`openrouter`	`OPENROUTER_API_KEY`	`https://openrouter.ai/api/v1`
`groq`	`GROQ_API_KEY`	`https://api.groq.com/openai/v1`
`google`	`GOOGLE_API_KEY`	`https://generativelanguage.googleapis.com/v1beta/openai/`
`openai`	`OPENAI_API_KEY`	`https://api.openai.com/v1`
`anthropic`	`ANTHROPIC_API_KEY`	`https://api.anthropic.com/v1/`
`litellm`	`LITELLM_API_KEY`	`LITELLM_BASE_URL`
`modal`	`MODAL_PROXY_TOKEN_ID` and `MODAL_PROXY_TOKEN_SECRET`	`MODAL_BASE_URL`
`local`	`REHEARSAL_LOCAL_API_KEY`	`REHEARSAL_LOCAL_BASE_URL`
`custom`	`REHEARSAL_CUSTOM_API_KEY`	`REHEARSAL_CUSTOM_BASE_URL`

local is for a model server on your own machine, such as vLLM. custom is for any other OpenAI-compatible address. rehearsal prices shows the price the server uses for each configured model.

Workers and sandboxes

Setting	Default	What it changes
`REHEARSAL_RUN_PARALLEL`	`4`	Into how many parts a run is cut. Each part needs one free worker and one sandbox.
`REHEARSAL_BUILD_PARALLEL`	`4`	How many jobs a build writes and validates at once. Each one being validated runs its own copy of the world.
`REHEARSAL_MAX_ACTIVE_BUILDS`	`0`, no limit	Builds that one project may have in progress.
`REHEARSAL_MAX_ACTIVE_RUNS`	`0`, no limit	Runs that one project may have in progress.
`REHEARSAL_CONTAINER_MEMORY`	`1g`	The memory limit of each container in a sandbox.
`REHEARSAL_CONTAINER_CPUS`	`1.0`	The processor limit of each container in a sandbox.
`REHEARSAL_CONTAINER_PIDS`	`256`	The process limit of each container in a sandbox.
`REHEARSAL_LEASE_SECONDS`	`60`	How long a worker's claim on a server job lasts without a renewal. After that, another worker takes the job.
`REHEARSAL_WORKER_CONTAINER`		The worker's own container id, when the worker runs in a container. The Compose file sets it.

Spend and plans

Setting	Default	What it changes
`REHEARSAL_MONTHLY_ALLOWANCE_USD`	`0`, no limit	A model spend limit for the whole server, for a calendar month. It applies when plans are off.
`REHEARSAL_PLANS`	`off`	`on` puts every workspace on a plan and enforces its limits. Leave it off on a private server.
`REHEARSAL_QUIET_HOURS_UTC`	`18:00-02:00`	When evaluations of the Free plan may start, as `HH:MM-HH:MM` in UTC. It may cross midnight.
`DODO_PAYMENTS_API_KEY`		The payment provider's key. Only with plans on.
`DODO_PAYMENTS_WEBHOOK_KEY`		The secret that signs the payment provider's notifications.
`DODO_PAYMENTS_ENVIRONMENT`	`test_mode`	`test_mode` or `live_mode`.
`DODO_PRODUCT_*`		The payment provider's product id for each thing that can be bought: `PRO_MONTHLY`, `PRO_YEARLY`, `TEAM_MONTHLY`, `TEAM_YEARLY`, `EPISODES_1000`, `BUILD_1`.

Tracing and training

Setting	Default	What it changes
`NEATLOGS_API_KEY`		With it, the server sends traces of model calls to neatlogs. Without it, tracing is off.
`KAGGLE_API_TOKEN`		For training, which is not built yet.

Configuration

ClientCopy Anchor Link

Database and filesCopy Anchor Link

Sign-inCopy Anchor Link

ModelsCopy Anchor Link

Model chainsCopy Anchor Link

ProvidersCopy Anchor Link

Workers and sandboxesCopy Anchor Link

Spend and plansCopy Anchor Link

Tracing and trainingCopy Anchor Link

On this page