{
  "markdown": "# Crowdcast\n\n**Multi-agent social simulation as a Claude Code skill.**\n\nDrop in a document, describe what you want to predict — Crowdcast spawns dozens of AI agents that argue, post, react, and evolve on a simulated platform. Then it writes you a report.\n\nNo APIs to configure. No databases to run. No web server. Just `/crowdcast simulate`.\n\n## Why Crowdcast over MiroFish?\n\n[MiroFish](https://github.com/666ghj/MiroFish) is the original multi-agent prediction engine that inspired this project. It's a great tool — but it comes with friction:\n\n| | MiroFish | Crowdcast |\n|---|---------|-----------|\n| **Setup** | Python + Node.js + Docker + env vars | `git clone` into `~/.claude/skills/` |\n| **External APIs** | Zep Cloud ($25/mo) + LLM API ($5-15/run) | None — included in Claude subscription |\n| **Rate limits** | Zep free tier: 1 simulation/month | Unlimited |\n| **Infrastructure** | Flask server + Vue.js frontend + Zep + OASIS | Zero — runs inside Claude Code |\n| **Resume after crash** | Manual — restart from scratch | `/crowdcast resume` picks up where it stopped |\n| **Agent scale** | Up to millions (via OASIS) | Up to ~100 (practical limit) |\n| **Agent depth** | Uniform — all agents use same LLM prompt | Hybrid — key agents think deeply, crowd is batched |\n| **Creative mode** | Limited (primarily forecast-oriented) | Full narrative simulation with character arcs |\n| **Interview agents** | Requires running server | `/crowdcast interview` — instant, in terminal |\n\n**TL;DR:** MiroFish is more powerful at massive scale. Crowdcast is simpler, cheaper, and more accessible for most use cases. If you need to simulate a million agents, use MiroFish. If you need a quick prediction from 50 agents with zero setup, use Crowdcast.\n\n## What It Does\n\nUpload a document (news article, research report, novel) and describe what you want to simulate:\n\n1. **Analyze** — extract entities, relationships, and context from your documents\n2. **Profile** — create AI agents with distinct personalities, stances, and behaviors\n3. **Simulate** — run rounds of agent interaction on a simulated platform\n4. **Report** — produce an analytical prediction or narrative retelling\n\n## Two Modes\n\n| | Forecast | Creative |\n|---|---------|----------|\n| **Input** | News, reports, policy docs | Fiction, scripts, scenarios |\n| **Agents** | Stakeholders + social groups | Characters + background |\n| **Simulation** | Social media (posts, comments, likes) | Free-form world (dialogue, actions, events) |\n| **Output** | Analytical prediction with trends | Narrative retelling with character arcs |\n\nMode is auto-detected from your document, or set explicitly with `--mode=forecast` / `--mode=creative`.\n\n## Installation\n\n**Option A: Clone directly into skills (simplest)**\n\n```bash\ngit clone https://github.com/TheQmaks/crowdcast.git ~/.claude/skills/crowdcast\n```\n\n**Option B: Symlink (keep repo separate)**\n\n```bash\ngit clone https://github.com/TheQmaks/crowdcast.git\nln -s $(pwd)/crowdcast ~/.claude/skills/crowdcast\n```\n\n**Option C: Install as Claude Code plugin (via marketplace)**\n\nThis repo doubles as a single-plugin Claude Code marketplace. Add it once, then install:\n\n```bash\n/plugin marketplace add TheQmaks/crowdcast\n/plugin install crowdcast@theqmaks\n```\n\nThat's it. No `npm install`, no `pip install`, no Docker, no `.env` files.\n\nVerify installation — type `/crowdcast` in Claude Code and you should see the help menu.\n\n## Usage\n\n```bash\n# Full simulation — forecast mode\n/crowdcast simulate ./news_report.pdf \"How will the public react to this policy change?\"\n\n# Full simulation — creative mode\n/crowdcast simulate ./chapter1.txt ./chapter2.txt \"Continue the story with these characters\"\n\n# Analyze documents only (no simulation)\n/crowdcast analyze ./report.pdf\n\n# Resume an interrupted simulation\n/crowdcast resume sim_a3f8b2c91d04\n\n# Regenerate report from completed simulation\n/crowdcast report sim_a3f8b2c91d04\n\n# Interview a simulated agent in character\n/crowdcast interview sim_a3f8b2c91d04 mayor_chen\n```\n\n## How It Works\n\n```\nDocument → [Analyzer] → Knowledge Graph → [Profiler] → Agent Personas\n                                                            ↓\nReport ← [Reporter] ← Round Logs ← [Simulator] ← Personas + Config\n```\n\n- **Claude is everything** — the LLM, the NER engine, the simulator, and the report writer\n- **Subagent orchestration** — each phase runs as an isolated Claude Code subagent with clean context\n- **File-based state** — all data stored as JSON in `.crowdcast/simulations/`\n- **Resumable** — every phase saves progress; interrupted simulations continue from the last checkpoint\n- **Hybrid depth** — key agents (leaders, influencers) think individually; crowd agents are batched for efficiency\n\n## Typical Scale\n\n| Parameter | Forecast | Creative |\n|-----------|----------|----------|\n| Agents | 50+ (8-10 key + crowd groups) | 5-20 characters |\n| Rounds | 50-100 | 30-50 |\n| Wall time | 20-40 min | 15-30 min |\n| Cost | Included in Claude subscription | Included in Claude subscription |\n\n## Requirements\n\n- [Claude Code](https://claude.ai/claude-code) (CLI, desktop app, or IDE extension)\n- Claude subscription (Pro, Team, or Enterprise)\n\n## Project Structure\n\n```\nSKILL.md                              # Main orchestrator — loaded when you type /crowdcast\nreferences/\n  data-schemas.md                     # All JSON schema definitions\n  phase1-analyzer.md                  # Document analysis subagent prompt\n  phase2-profiler.md                  # Persona generation subagent prompt\n  phase3-simulator-forecast.md        # Forecast simulation subagent prompt\n  phase3-simulator-creative.md        # Creative simulation subagent prompt\n  phase4-reporter.md                  # Report generation subagent prompt\ntest_seeds/\n  sample_news.md                      # Sample document for testing\n```\n\n## Inspired By\n\n[MiroFish](https://github.com/666ghj/MiroFish) by the MiroFish team at Shanda Group, powered by [OASIS](https://github.com/camel-ai/oasis). Crowdcast reimagines the concept as a zero-dependency Claude Code skill.\n\n## License\n\nMIT\n",
  "bytes": 6081,
  "sha": "7e647be92a4e63b59ab0fa9af71043ab5e9cd6c00f4f3e52b7364751173c84a8",
  "repo_slug": "theqmaks/crowdcast",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_theqmaks_crowdcast_crowdcast_4e32aaba/readme"
}