Minerva takes a raw idea and drives it to a finished, human-approved spec — an epic with dependency-tracked stories — over a stable subprocess ABI. No terminal session to babysit, no UI of its own. Free and open source.
$ curl -fsSL https://mdostal.github.io/minerva/install.sh | bash
Installs the CLI, wires the MCP server into whatever agent harness is already on the
machine (Claude Code, Codex CLI today), and drops in the usage skill so an agent gets the
full startRun → poll → answer → repeat pattern immediately. Already
installed? minerva agent init is the same thing, idempotent.
Getting from a raw idea to an approved plan meant a human hand-running
/hive:kickoff and /hive:plan in a terminal. Minerva makes that
step a service you can drive from anywhere: feed it an idea over a stable
subprocess ABI, and it drives plugin-hive's kickoff and plan skills against an isolated
per-run git workspace, surfaces exactly the questions that need a human answer, and
returns the finished spec.
Splitting the planner out of the terminal is what lets you feed many ideas at once — each idea gets its own run and its own decision thread, in parallel, instead of a serial per-idea session. There is no daemon; nothing advances a run on its own. A run only moves when a caller answers its questions.
Minerva plans; it doesn't route and it doesn't provision. Ideas arrive already routed, questions go to a human, and the approved output moves on to build the thing.
Minerva's public interface is the Pantheon subprocess ABI itself —
{method, params} in, {result} or {error} out — wire
compatible with plugin-hive's task-tracking adapter ABI. The same eight methods run as an
MCP server (minerva mcp) for anything that speaks MCP.