founder-os

Multi-agent primitives for Claude Code.

Claude Code gives you one answer. founder-os lets you ask multiple agents, compare consensus, surface disagreement, and synthesize the useful part.

claude plugin install rhinehart514/founder-os

Try this first:

/founder-os:stochastic n=10 Should we use Postgres or SQLite for this project?

It polls 10 agents with the same prompt, then aggregates consensus, divergences, and outliers.

The primitives

Skill	What it does
/stochastic	Poll N agents (default 10) with identical prompts. Aggregate by consensus, divergences, outliers.
/model-chat	Spawn 5 agents into a debate room. Round-robin turns, shared history, synthesizer merges.
/fanout	Fan-out N researchers (Sonnet) in parallel, fan-in with Opus synthesis.
/skillbuilder	Build flawless Claude Code skills. Studies existing skills as reference, ensures correct format, pushes for genuine intelligence.
/autoresearch	Autonomous hill-climb. Read code, generate mutations, measure, keep-or-revert, repeat.

Install

Plugin install

claude plugin install rhinehart514/founder-os

Skills become /founder-os:stochastic, /founder-os:model-chat, /founder-os:fanout, /founder-os:skillbuilder, /founder-os:autoresearch.

Direct install (short skill names)

npm i -g @rhinehart514/founder-os

Or from source:

git clone https://github.com/rhinehart514/founder-os.git
cd founder-os && bash bin/install.sh

Skills become /stochastic, /model-chat, /fanout, /skillbuilder, /autoresearch.

Restart Claude Code after installing.

Usage

/stochastic Should we use Postgres or SQLite for this use case?
/stochastic n=5 What's the best approach to rate limiting here?

/model-chat Is this API design good?
/model-chat agents=3 rounds=3 Should we ship this feature?

/fanout What are the tradeoffs of server components vs client components?
/fanout n=3 model=haiku Research authentication patterns for this stack

/skillbuilder Create a skill for code review with verification gates
/skillbuilder audit /stochastic

/autoresearch Optimize the Lighthouse score
/autoresearch Shrink the bundle size

Why this works

Same model, different runs = systematically different outputs. Polling multiple agents filters hallucinations (consensus), surfaces genuine judgment calls (divergences), and catches ideas a single pass misses (outliers).

What's included

skills/
  stochastic/SKILL.md
  model-chat/SKILL.md
  fanout/SKILL.md
  skillbuilder/SKILL.md
  autoresearch/SKILL.md
  autoresearch/templates/   # hill-climb runner, dashboard, config
agents/
  parent.md      # orchestrator
  qa.md          # quality checker
  researcher.md  # research agent

License

MIT