| name | rl-env-use-install |
|---|---|
| description | Install rl-env-use into the current agent (Claude Code, Codex, Hermes, Openclaw, etc.) and set up the Browser Use and OpenRouter API keys so the user can build an environment right away. |
Use this file only for first-time install or reconnect. For daily use, read SKILL.md.
You're setting up a tool that turns one line ("predict whether a clinical trial met its primary endpoint from its
registered protocol") into an RL environment built from real cases that Browser Use agents mine from the web, and then
hillclimbs a model on it. After install the user says what skill they want an environment for; you run it by reading
SKILL.md.
Four things must exist on this machine:
- The
rl-env-userepo cloned somewhere stable, installed into its own virtual environment (Python 3.11+). - A Browser Use API key (cloud browsers for the mining agents) and an OpenRouter API key (every LLM call: planner,
task writer, critic, frontier panel, the models under test) in
.envat the repo root. - Optionally Node.js (>= 18) on
$PATH: it writes the HTML reports. Without it everything works and the reports are skipped. - The current agent can discover
SKILL.md(a skills directory, or an import in itsCLAUDE.md-equivalent).
- Do everything yourself. Only ask the user for the two API keys.
- Prefer a stable clone path like
~/Developer/rl-env-use(not/tmp, not~/Downloads). - Never spend credits during install. Every check below is free.
test -d ~/Developer/rl-env-use || git clone https://github.com/browser-use/rl-env-use ~/Developer/rl-env-use
cd ~/Developer/rl-env-use
command -v uv >/dev/null && uv venv && uv pip install -e '.[test]' || (python3 -m venv .venv && .venv/bin/pip install -e '.[test]')
.venv/bin/python -m pytest -q # offline, no keys needed, ~10 s
If the repo is already there, git pull --ff-only and re-run the install line.
- Claude Code:
mkdir -p ~/.claude/skills && ln -sfn ~/Developer/rl-env-use ~/.claude/skills/rl-env-use - Codex:
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills" && ln -sfn ~/Developer/rl-env-use "${CODEX_HOME:-$HOME/.codex}/skills/rl-env-use" - Another agent: symlink the repo into its skills directory as
rl-env-use, or point its system prompt at~/Developer/rl-env-use/SKILL.md.
Put the CLI on the user's path so it runs from any folder: ln -sfn ~/Developer/rl-env-use/.venv/bin/rl-env-use ~/.local/bin/rl-env-use
(any directory on $PATH works).
Check what is already there and stop at the first hit for each key: an exported BROWSER_USE_API_KEY /
OPENROUTER_API_KEY, or a non-empty line in ~/Developer/rl-env-use/.env. For each missing key, ask the user once:
I need a Browser Use API key for the mining agents' cloud browsers - get one at https://cloud.browser-use.com - and an OpenRouter API key for the language models - https://openrouter.ai/keys. Paste them here and I'll write them to
~/Developer/rl-env-use/.env.
cp -n .env.example .env && chmod 600 .env # then fill in the two values; never echo them back, never commit .env
Free checks:
.venv/bin/python -c "from rl_env_use.llm import load_dotenv, bu_balance; load_dotenv(); print('Browser Use credits:', bu_balance())"
curl -s -o /dev/null -w '%{http_code}\n' -H "Authorization: Bearer ${OPENROUTER_API_KEY:-$(sed -n 's/^OPENROUTER_API_KEY=//p' .env)}" https://openrouter.ai/api/v1/key
None or 401 means a wrong key: ask once more.
.venv/bin/rl-env-use run envs/trial-forecast --model qwen-35b --flow models --dry-run # prices a run from the shipped env's measurements; writes nothing
Tell the user, in one short message: where it is installed, that a good first message is "build me an RL environment
for from ", that a build costs real credits (the shipped envs cost $20-75 each; create
stops at its --budget, $40 by default), and that everything lands in ./envs/<name>/.
cd ~/Developer/rl-env-use && git pull --ff-only && uv pip install -e . (or .venv/bin/pip install -e .).