.env.example

# Copy to `.env` (git-ignored) and fill in. `docker run` does not auto-load it: it is passed
# explicitly with `--env-file .env` (see CLAUDE.md) and by the dev container's runArgs. The GPU
# `make` targets pass it via ENV_FILE; the CPU targets pass no credentials at all. `docker compose`
# is the exception — it reads a repo-root `.env` on its own, for `${VAR}` interpolation.
#
# Leave a key blank rather than deleting the line: the toolkit reads a set-but-empty value as unset.

# Weights & Biases run logging (src/training/environment.py).
WANDB_API_KEY=

# Gated HuggingFace model/dataset access.
HF_TOKEN=

# Region for the optional S3 dataset sources — bring your own AWS credentials
# (agent-docs/infrastructure/aws-auth.md).
AWS_DEFAULT_REGION=

# External LLM + web-search backends for the RL environments, all optional
# (agent-docs/training-methods/grpo/environments/native-tool-use.md).
OPENROUTER_API_KEY=
OPENAI_API_KEY=
TAVILY_API_KEY=
SERPER_API_KEY=
BRAVE_API_KEY=

# Rollout servers on other nodes (agent-docs/infrastructure/rollout-servers.md): the trainer address
# the server's workers dial back to for the weight-sync group, and the one address the group's
# rendezvous store listens on (every interface under `HALO_WEIGHT_SYNC_BIND_ALL=1`), when the
# default-route interface is not one the server can reach. IPv4 only.
# Unset = auto (loopback for a same-host server, else the default route).
VLLM_GROUP_HOST=
SGLANG_GROUP_HOST=