Roast your startup with 500 AI users in 60 seconds.
Founder validation through AI swarm simulation. Paste a pitch and watch 100–500 AI users react like a real Reddit / HN / Product Hunt thread — objections, questions, snark, and silence — then walk away with a decision brief: ship, sharpen, or kill, plus the exact questions to ask real users next.
👉 Try it live: swarmie.vercel.app
Pitch in → 500 AI users react → insights out. 14s demo · download mp4
Before you spend $10k on user interviews or burn 6 months on the wrong positioning, find your top 3 objections in 60 seconds.
Swarmie isn't a replacement for talking to real users. It's a pre-interview filter — kill 17 bad positionings before you embarrass yourself with the 18th.
Built for:
- Pre-launch founders validating a pitch
- PMs A/B testing positioning before a launch
- Marketers stress-testing landing-page copy
- Indie hackers without a panel of real users to call
- Paste your pitch (problem / product / audience / pricing / competitors).
- Swarmie parses it, then builds a swarm of 100–500 agents with distinct personas, biases, and tone.
- Agents react: post, comment, upvote, ask questions, raise objections — or scroll past. Each persona rolls its action from its archetype's likelihood distribution; silent ignores and upvotes cost zero tokens.
- You get a decision brief, not a vanity score:
- Verdict —
ship it/sharpen/wrong audience/kill, with a confidence band. - Next action — the single most important move before writing more code.
- Top objections — each with the exact question to ask 5 real users, a kill-criteria, and a suggested fix.
- Why they scrolled past — clustered, decision-useful reasons the silent majority ignored you.
- Supporting signal: sentiment split, per-ICP fit, messaging gaps.
- Verdict —
- Click any agent to read its persona and chat with it. Thumbs-up/down each objection to train the swarm.
Powered by:
- A lean, in-process async LLM swarm (no heavy simulation deps) with a hard cost ceiling.
- Two-tier model routing (cheap reactions / deep agents + synthesis) and a transparent Gemini fallback.
- Any OpenAI-compatible LLM (OpenAI, Anthropic via OpenRouter, Groq, Together, DeepSeek, Ollama for local/free).
- Supabase for privacy-light analytics, run history, and feedback (optional).
| Tool | Version | Check |
|---|---|---|
| Node.js | 18+ | node -v |
| Python | 3.11 – 3.12 | python --version |
| uv | latest | uv --version |
# 1. Clone
git clone https://github.com/hp-8/swarmie.git
cd swarmie
# 2. Configure
cp .env.example .env
# edit .env — add your LLM_API_KEY (required).
# Optional: LLM_FALLBACK_API_KEY (Gemini fallback), ROAST_MAX_COST_USD (cost cap),
# VITE_SUPABASE_URL + VITE_SUPABASE_ANON_KEY (analytics + feedback).
# Leave LLM_API_KEY pointed at a local Ollama to run for $0.
# 3. Install everything
npm run setup:all
# 4. Run
npm run devOpen http://localhost:3000. Backend runs on :5001.
cp .env.example .env
docker compose up -d🚧 Alpha. Forked from MiroFish — a Chinese-language swarm-prediction engine — and pivoted toward founder validation, then rebuilt on a slim in-process LLM swarm (the heavy OASIS/Zep stack is legacy, import-guarded, and no longer required).
Shipped: end-to-end 60s pipeline · decision-brief output (verdict + per-objection user-tests + kill-criteria) · "why they scrolled past" silence analysis · per-agent chat · two feedback layers (per-objection votes + product feedback) feeding calibration · atmospheric, fully-responsive marketing site · PDF export · Supabase analytics.
In development: live grounding in real Reddit/HN/review chatter, public backtest calibration, and the paid tiers (live voice interviews, continuous monitoring, real-user bridge).
See ROADMAP.md for what's coming.
Swarmie keeps MiroFish's strong sim core (OASIS + Zep + multi-step pipeline) but pivots:
| Dimension | MiroFish | Swarmie |
|---|---|---|
| Audience | General prediction (news, novels, policy) | Pre-launch founders |
| Input | Any seed document | Pitch / deck / one-pager |
| Output | Long-form prediction report | Decision brief: verdict + objection user-tests + silence analysis |
| Realism | LLM personas from raw graph | Sampled, cited silence reasons today; live grounding (WIP) |
| Calibration | None | Public backtest scoreboard (WIP) |
| Language | Chinese-first | English-first |
| Cost focus | Frontier models | Tiered routing + local-model default |
We need help on:
- ICP corpora — scrape + tag Reddit/HN comments by founder-relevant segments
- Backtest cases — pick known startups (hits + flops), measure prediction accuracy
- Cost optimization — tiered model routing, archetype clustering, prompt caching
- Founder UX — make the "upload pitch → see report" flow brutally simple
See CONTRIBUTING.md.
AGPL-3.0. See LICENSE.
Inherited from upstream MiroFish. Network-use clause applies — if you host Swarmie as a service, you must publish your modifications.
For commercial licensing or hosted-service exceptions, open an issue.
Swarmie stands on the work of:
- MiroFish by 666ghj and team — upstream codebase
- OASIS by CAMEL-AI — agent social simulation engine (Apache 2.0)
- Zep — agent memory + knowledge graph
- Shanda Group — strategic support for original MiroFish
See NOTICE for full attributions.