Run RLE-Bench
Install, evaluate, and inspect results with the RLE-Bench CLI.
01Install
We test RLE-Bench on Linux with Git, Make, uv, and Docker. GPU tasks also require an NVIDIA GPU accessible from Docker.
git clone https://github.com/RLE-Bench/RLE-Bench.git
cd RLE-Bench
make install
source .venv/bin/activate02Check your setup
List available tasks and check for missing dependencies.
rlebench list
rlebench doctor03Run an evaluation
Prepare Mobile Base Design assets and images. No GPU required.
rlebench prepare task08Choose your agent and credentials below. Replace MODEL_NAME with a model available to your account.
Anthropic · Subscription
Use a Claude Code subscription OAuth token, supplied as CLAUDE_CODE_OAUTH_TOKEN. This is different from ANTHROPIC_AUTH_TOKEN.
export CLAUDE_CODE_OAUTH_TOKEN="YOUR_OAUTH_TOKEN"rlebench run task08 -a claude-code \
-e anthropic/claude-code -m "anthropic/MODEL_NAME"Anthropic · API
Use an Anthropic API key. ANTHROPIC_AUTH_TOKEN is also supported by this API configuration; use one credential variable.
export ANTHROPIC_API_KEY="YOUR_API_KEY"rlebench run task08 -a claude-code \
-e anthropic/api -m "anthropic/MODEL_NAME"Codex · Subscription
Sign in with ChatGPT using file-based credential storage. The benchmark uses ~/.codex/auth.json.
codex -c 'cli_auth_credentials_store="file"' login
export CODEX_FORCE_AUTH_JSON=1rlebench run task08 -a codex \
-e openai/codex -m "openai/MODEL_NAME"Codex · API
Run the Codex agent with an OpenAI API key.
export OPENAI_API_KEY="YOUR_API_KEY"rlebench run task08 -a codex \
-e openai/api -m "openai/MODEL_NAME"Gemini · API
Use GEMINI_API_KEY to run Gemini on task09. Prepare the task first with rlebench prepare task09.
export GEMINI_API_KEY="YOUR_API_KEY"rlebench run task09 -a agy \
-e gemini/api -m gemini-3.7-flashGLM · API with Claude Code
Run GLM through Claude Code using ZAI_API_KEY (or GLM_API_KEY). Use zai/api-cn for the China endpoint.
export ZAI_API_KEY="YOUR_API_KEY"rlebench run task08 -a claude-code \
-e zai/api -m "zai/MODEL_NAME"GPU example
Prepare the Interactive Control task family.
rlebench prepare task01With Anthropic API credentials configured, run the open-fridge task on GPU 0.
rlebench run task01/L1/01-open-fridge \
-a claude-code -e anthropic/api -m "anthropic/MODEL_NAME" --device cuda:0Use rlebench run --help for agents and options. Check task docs for required datasets and weights.
04Bring your own agent
You can use any Harbor-supported agent and model. Choose the agent with -a and the model with -m, then add Harbor configuration after --. RLE-Bench forwards everything after -- directly to Harbor.
rlebench run task08 -a AGENT_NAME -m "PROVIDER/MODEL_NAME" \
-- --ak "OPTION=VALUE"Replace the placeholders with your agent, model, and supported agent options. For agents without a built-in endpoint configuration, omit -e and supply credentials and allowed API hosts through Harbor as needed. Run harbor run --help for available options.
05View results
Generate reports with run statuses, rewards, and costs.
rlebench summarize jobsOpen the viewer to explore trajectories, scores, and recordings.
rlebench view jobs