Running in about two minutes.
Leverage needs one source of intelligence and one execution fabric. The fastest path uses something you already have and asks for no API key at all.
Install
git clone <your fork> leverage cd leverage npm install cp .env.example .env.local
Connect intelligence
At least one of these, ordered by least setup. The first two need no key, because your subscription already paid for the model.
| Source | Key needed | What it is |
|---|---|---|
| Your agent CLI | none | Already installed and logged in: claude, codex, gemini, opencode. |
| Your MCP host seat | none | Run Leverage inside your agent; it samples the host model. |
| Ollama | none | Every model you have pulled. Nothing leaves the machine. |
| OpenAI-compatible | yours | LM Studio, vLLM, llama.cpp, a gateway of your own. |
Leverage detects agent CLIs on PATH and probes whether each is signed in. One that is installed but signed out is excluded from the auction with the command that fixes it, rather than hired and left to fail.
Connect the execution fabric
RocketRide runs the worker pipelines. Sign in through its CLI so no key passes through your clipboard:
pnpm exec rocketride login # writes ROCKETRIDE_APIKEY into .env npm run verify:rocketride # proves the whole path end to end
The endpoint is https://staging.rocketride.ai. Note that the published SDK docs name a different host; the running system disagrees with them in three places, all recorded in docs/ROCKETRIDE_FINDINGS.md.
Run a mission
npm run probe:models # measure what your models can actually do npm run mission # the benchmark mission, for real npm run dev # Mission Control at http://localhost:3000
The probe is worth running first. It costs a couple of minutes and stops the auction hiring a model that returns an empty response to every structured request. That is what happened here on the first real run.