A Claude Code plugin that runs MiroFish-style social simulations. Provide seed material (articles, scenarios, prediction questions), and the plugin generates agent personas, runs multi-round simulations with independent subagents, and produces HTML analysis reports.
- Init — Extract stakeholders from seed material, generate rich personas (~2000 words each), map relationships
- Run — Each round: activate agents by probability, batch by influence (early/mid/late), dispatch one independent subagent per agent per batch in parallel
- Report — Analyze emergent dynamics, interview pivotal agents, generate self-contained HTML report
Each agent is a stateless, independent LLM call — matching the MiroFish/OASIS architecture. A JSON state file serves as external memory (replacing Zep/Neo4j). Agents see only their persona, opinion history, and a filtered feed of recent posts.
- 1 fresh subagent per persona per round — no single-agent role-playing shortcuts
- 3 influence-weighted batches per round — high-influence agents speak first, setting the tone
- JSON state as external memory — read before each round, written after all batches complete
- PreToolUse hook — blocks background Agent calls that would fall back to single-agent mode
# Add the marketplace
claude plugins marketplace add madiha-right/hivemind
# Install the plugin
claude plugins install hivemind@hivemind/hivemind
Then provide seed material (paste text, reference a file, or describe a scenario). The plugin handles persona generation, simulation, and reporting.
.claude-plugin/
plugin.json # Plugin manifest
marketplace.json # Marketplace config
skills/
hivemind/
SKILL.md # Main orchestrator skill
references/
persona-guide.md # Persona generation template
report-guide.md # HTML report generation guide
agents/
hivemind-agent.md # Single-agent-per-round definition
hooks/
hooks.json # PreToolUse hook config
check-hivemind-agent # Blocks background Agent calls
run-hook.cmd # Cross-platform hook runner