41 callable skills for any app that talks to humans. Emotion scan, crisis detection, coaching, community engagement, wellness scoring — all over P2P sidechannels on Trac Network's Intercom. No REST endpoints, no API keys, no cloud dependency.

Your app already does the job. 5FAN makes it feel right — so notifications land with the right tone, crisis signals never get missed, and responses sound like someone who actually listened.

Live Demo · Architecture · Skill Guide · Fork Guide

trac1wtsn8ru2ryknk36rd6glp2tfj0dawnh2gkjrg90gzqvwes65v78qmwjuzq

• Live Demo: https://joeatang.github.io/5FAN/
• Production VPS: Running 24/7 on Trac Network (peer + skill server on 104.131.183.26)
• SC-Bridge: WebSocket skill invocation active — connected: true, authed: true, healthy: true

Call any skill over WebSocket or P2P sidechannel. Get a structured result back in milliseconds.

{ "type": "skill-call", "skill": "emotion-scan", "input": { "text": "I feel lost" } }
→ { "type": "skill-result", "result": { "ok": true, "signal": 0.82, "emotions": ["sadness", "confusion"], "category": "pain" }, "ms": 12 }

Every skill is deterministic first (keyword scanning, rule-based analysis), LLM-enriched second (local → cloud → template fallback). Crisis detection is always deterministic — no probabilistic misses.

All 41 skills are registered in skill-protocol.js and dispatched via skill-dispatch.js. Rate limited to 30 calls/min/caller.

{ "type": "skill:call", "skill": "compass-locate", "input": { "text": "I don't know what I want" } }
→ { "type": "skill:result", ... }

{ "type": "skill-call", "skill": "wellness-score", "input": { "text": "I've been off for days" } }
→ { "type": "skill-result", "skill": "wellness-score", "result": { "ok": true, ... }, "ms": 18 }

For simple cases, you should. But 5FAN solves three problems a prompted LLM can't:

Under the hood, 5FAN runs five specialized brains in parallel on every message. A consensus engine synthesizes their signals into one informed response.

View's curateConsensus() ranks signals, identifies the dominant brain, and builds a synthesis prompt that enriches the LLM system message. Crisis detection bypasses all of this — Hear's keyword scanner fires immediately and provides hotline resources with no LLM delay.

User message
    │
    ├─→ Hear.scan()    ─→ { signal: 0.7, emotions: ['frustration'], category: 'pain' }
    ├─→ Inspyre.scan() ─→ { signal: 0.4, themes: ['growth'] }
    ├─→ Flow.scan()    ─→ { signal: 0.2, category: 'recovery' }
    ├─→ You.scan()     ─→ { signal: 0.3, markers: ['self-reflection'] }
    └─→ View.scan()    ─→ { signal: 0.5, category: 'perspective' }
                               │
                  View.curateConsensus()
                               │
                               ▼
               { dominantBrain: 'hear',
                 synthesisPrompt: '...',
                 activeBrainCount: 4 }
                               │
                    ┌──────────┴──────────┐
                    │  LLM Bridge         │
                    │  local → cloud →    │
                    │  template fallback  │
                    └──────────┬──────────┘
                               │
                               ▼
                        Final response

Any agent on Intercom can invoke a single brain or the full swarm over sidechannels:

• Node.js 22.x or 23.x
• Pear Runtime
• LLM (optional): LM Studio or Ollama for local; Groq/OpenRouter/Together.ai for cloud

git clone https://github.com/joeatang/5FAN.git
cd 5FAN
npm install

# Terminal 1 — Admin peer
pear run . --peer-store-name admin --msb-store-name admin_msb \
  --dht-bootstrap "node1.hyperdht.org:49737,node2.hyperdht.org:49737"

# Terminal 2 — Second peer
pear run . --peer-store-name peer2 --msb-store-name peer2_msb \
  --dht-bootstrap "node1.hyperdht.org:49737,node2.hyperdht.org:49737"

# In both terminals
join 0000intercom
send Hello world

5FAN scans the message across all five brains, builds consensus, enriches via LLM (if available), and broadcasts a response.

Start LM Studio or Ollama — 5FAN auto-detects at localhost:1234 (LM Studio) or localhost:11434 (Ollama).

# LM Studio: just start the application and load a model
# Ollama:
ollama pull llama3.2:3b && ollama serve

Set environment variables in your shell or .env:

export FIVEFAN_LM_KEY=gsk_your_groq_key_here
export FIVEFAN_LM_CLOUD_URL=https://api.groq.com/openai
export FIVEFAN_LM_CLOUD_MODEL=llama-3.3-70b-versatile

Supported providers (all OpenAI-compatible):

• Groq — https://api.groq.com/openai
• OpenRouter — https://openrouter.ai/api
• Together.ai — https://api.together.xyz

Works fine without any LLM — falls back to curated templates (40+ per brain, ~200 total).

Edit config.js to toggle capabilities:

features: {
  feedReplies:   true,   // Auto-reply to community posts
  trainer:       true,   // 1:1 conversation mode
  nudges:        true,   // Contextual nudges in responses
  proactive:     true,   // Scheduled community posts
  crossBrain:    true,   // Multi-brain consensus
  userProfiling: true,   // Per-user word-frequency tracking
}

Kill switch: Set FIVE_FAN.enabled = false to disable all brain responses instantly.

DM the agent to start a 1:1 session:

Timezone-aware scheduled posts to the community channel:

10 templates per slot, LLM-enriched when available. Configurable timezone in config.js.

When using Express routes (server/routes.js):

5FAN/
├── brains/
│   ├── 5fan.js              # Shared config, constants, helpers
│   ├── hear/                # Emotional scanner + skill.json
│   ├── inspyre/             # Values alignment + skill.json
│   ├── flow/                # Habit guardian + skill.json
│   ├── you/                 # Data analyst / profiler + skill.json
│   └── view/                # Curator + curateConsensus() + skill.json
├── skills/
│   ├── eq-engine/           # 9 skills: emotion-scan, crisis-detect, reframe...
│   │   └── data/            # Emotion families, micro-moves, desire map
│   ├── compass/             # 5 skills: locate, interpret, point, practice, shift
│   ├── community/           # 5 skills: feed-reply, proactive-post, pulse...
│   ├── coach/               # 10 skills: tone-match, coach-chat, nudge-engine...
│   └── internal/            # 6 skills: earn-calculator, tier-gate, anti-bot...
├── server/
│   ├── brain-swarm.js       # Parallel scan + consensus engine
│   ├── lm-bridge.js         # Multi-provider LLM (auto-fallback)
│   ├── skill-server.js      # 41-skill handler imports + registration
│   ├── feed-responder.js    # Community feed auto-reply
│   ├── proactive-scheduler.js
│   ├── trainer-api.js       # 1:1 conversation manager
│   └── routes.js            # Express REST API
├── skill-dispatch.js        # Central skill dispatcher (35 callable)
├── skill-http.js            # HTTP skill endpoint
├── skill-protocol.js        # Skill registry + message types + channels
├── config.js                # Master config + feature flags
├── app-context.js           # System prompt identity
├── user-profile.js          # Onboarding + profile persistence
├── intercom-swarm.js        # P2P brain swarm routing
├── lm-local.js              # Local LLM adapter
├── lm-cloud.js              # Cloud LLM adapter
├── index.js                 # Main Intercom entry point
├── tests/                   # Skill test suites
└── ARCHITECTURE.md          # Detailed architecture docs

Each brain directory contains three files:

• roleConfig.js — Title, personality description, trigger keywords, template responses
• functions.js — scan(), fulfill(), log(), sendTo()
• index.js — shouldRespond(), handleMessage() (threshold-based)

5FAN is designed to be forked. The architecture separates what the brains do from how the system works — so you change the personality without touching the engine.

# 1. Clone
git clone https://github.com/joeatang/5FAN.git my-brand-agent
cd my-brand-agent
npm install

# 2. Edit app-context.js — change the personality
#    "You are VeeFAN, a character-trait companion for the VeeFriends community..."

# 3. Edit brains/*/roleConfig.js — change keywords + templates
#    Hear: swap emotion words for your domain
#    Inspyre: swap motivational triggers
#    Flow: swap habit/consistency keywords
#    You: swap identity/data markers
#    View: swap synthesis templates

# 4. Edit brains/*/skill.json — update skill manifests for your brains
#    Change "encodes", "domain", "tagline", "whenToUse" for each brain

# 5. Edit config.js — your timezone, LLM provider, feature flags

# 6. Run it
pear run . --peer-store-name admin --msb-store-name admin_msb \
  --dht-bootstrap "node1.hyperdht.org:49737,node2.hyperdht.org:49737"

You now have a 5-brain AI agent on a P2P network with multi-provider LLM, 1:1 coaching, proactive community engagement, and crisis detection — customized for your brand.

Need a brain that doesn't exist? Add one:

1. mkdir brains/mybrain — create the directory
2. Copy roleConfig.js, functions.js, index.js from any existing brain
3. Change the scan keywords, templates, and personality
4. Create a skill.json manifest (copy from any brain, update metadata)
5. Register in brains/5fan.js → add to BRAINS array
6. Register in skill-protocol.js → add to SKILL_REGISTRY
7. Import in server/brain-swarm.js → add to scan array

View's curateConsensus() automatically incorporates the new brain's signals — no other changes needed.

Built by: @joeatang

Built on: Trac Network Intercom · Pear Runtime · Trac Network

LLM providers: LM Studio · Ollama · Groq · OpenRouter · Together.ai

Version: 2.1.0
License: Apache-2.0