quickstart.md

Quick Start

Back to README · See also evaluation.md, execution-modes.md, and release-snapshot.md

Install

pip install -e .
pip install -e ".[dev]"

Optional Google-backed provider dependencies for online Calendar, Gmail, or Tasks paths:

pip install -e ".[google]"

Generate the benchmark snapshot

python scripts/generate_tasks.py

By default this writes under openclaw_env/data/{tasks,datasets}.

Run a first hard-benchmark evaluation

python examples/train_and_eval.py \
  --agent llm \
  --llm-provider openai \
  --llm-base-url https://api.example.com/v1 \
  --model Kimi-K2.5 \
  --task-prefix hard_decision_workflow_ \
  --split test \
  --mode multi \
  --max-steps 20 \
  --llm-max-tokens 192 \
  --llm-timeout-s 45 \
  --llm-request-retries 4 \
  --llm-retry-backoff-s 4 \
  --inter-task-sleep 3 \
  -v \
  --verbose-log ./tmp/hard_test_verbose.log \
  --save-report ./tmp/hard_test_report.json

This uses the documented benchmark-facing protocol rather than the raw CLI defaults. The CLI default step budget is 15, while standard hard-benchmark runs in this repository typically use 20.

Run Claude through a local LiteLLM proxy

Export the proxy settings:

export LITELLM_PROXY_KEY="your-litellm-proxy-key"
export LITELLM_PROXY_BASE_URL="http://127.0.0.1:4000/v1"

Then run:

python examples/train_and_eval.py \
  --agent llm \
  --llm-provider openai \
  --model claude-sonnet-4.6 \
  --task-prefix hard_decision_workflow_ \
  --split test \
  --mode multi \
  --max-steps 20 \
  --llm-max-tokens 1024 \
  -v

When --llm-provider openai is used with a claude-* model name, the client defaults to the local LiteLLM proxy URL and reads LITELLM_PROXY_KEY unless --llm-base-url or --llm-api-key-env is overridden.

Common next steps

• evaluation.md: output interpretation and comparison commands
• execution-modes.md: exact mode behavior
• openclaw-setup.md: real and hybrid prerequisites
• task-generation.md: generator flags and release-profile overrides