@star-factory/launch-cli 0.4.0. Install with npm install --global @star-factory/launch-cli@0.4.0, or run one-off with npx -y @star-factory/launch-cli@0.4.0 <command>. Every command accepts --json and returns one stable envelope per command (schemaVersion: star.launch-agent/v2) carrying ok, command, the environment that answered, data, and error.
Environment
Production is the default and needs no selector.--env staging targets staging and --api-url <origin> targets a custom origin; both are global, may appear before or after the command, and apply to that one invocation only. Nothing is remembered between commands, so a staging session carries --env staging on every command (or exports STAR_ENV=staging). Tokens are stored per origin, so each environment is authorized separately.
Launch commands
Founder decisions
star launch create and star launch validate carry four founder decisions. None has a default: on a terminal create asks for whatever you left out; with --json, off a terminal, or from a file that omits one, it returns DECISION_REQUIRED naming the unanswered questions.
The tier ids, equity-link bounds, and the token amounts each choice reserves are published by
star launch requirements and returned in the --dry-run preview. The CLI ships no economics of its own.
Recommended sequence: launch requirements → launch login → launch create … --dry-run --json → review → launch create … --yes --json.
Rewards commands
MCP server
Every tool that talks to Star accepts a per-call
environment (production or staging) or apiUrl; without one it uses the environment star mcp serve was started with. Every result names the environment it answered for. A call that omits a founder decision is answered with DECISION_REQUIRED and the unanswered questions in error.details.questions, never a default.
Agents should parse structuredContent, not scrape text. Invalid arguments return INVALID_TOOL_INPUT before any network call.
Use it from Codex or Claude
mcp add registers the tools. It does not install the skill. Until the star-launch-agent plugin is listed, install the CLI, copy the skills, then register MCP. The same steps, with host-specific paths, are in the Quickstart.
Errors
Security model: agents never hold wallet secrets or keypairs, and nothing launches until you approve the exact payload. See the Quickstart.