CypherVibe

An autonomous AI C-suite that runs a micro-business end-to-end. Eight specialized agents — CEO, CMO, Engineer, CFO, Skeptic, Compliance, Anthropologist, Brand Steward — debate strategy with evidence gates and a Skeptic veto, then execute. The human only signs off on what's legally imprescindible (KYC, payments, contracts).

Try it in 60 seconds

git clone https://github.com/liberathio/cyphervibe.git && cd cyphervibe
pnpm install && pnpm db:up && pnpm db:migrate
pnpm council:demo   # 8 agents debate a synthetic objective, no API keys needed

You'll watch eleven stages run: intake → diverge → evidence-gate → steel-man → propose → adversarial-double → decide → gate-check → execute → verify → learn. Entirely offline; no LiteLLM, no Telegram, no money touched.

→ Architecture · Why it might work · Prior-art survey · v0 case study

Status

v0 — scaffolding. Documentation complete in docs/. Code skeletons for both councils, KPI gates, skill library, Telegram approval bot. No real money is touched in v0.

Architecture at a glance

                ┌─────────────────────────┐
   signals ───▶ │   DISCOVERY COUNCIL     │ ── monthly shortlist ──┐
   (Reddit/HN/  │   3 new + 4 shared      │                        │
    PH/GH…)     │   3-arbitrages scoring  │                        ▼
                └─────────────────────────┘                  ┌─────────┐
                                                             │ HUMAN   │
                ┌─────────────────────────┐                  │  picks  │
                │   EXECUTION COUNCIL     │ ◀─ objective ────┤ a niche │
   skill ──RAG─▶│   8 roles, 11 stages    │                  └─────────┘
   library      │   adversarial veto      │
   (pgvector)   └────────────┬────────────┘
                             │
                       ┌─────┴──────┐
                       ▼            ▼
                ┌──────────┐  ┌─────────────┐
                │ KPI GATE │  │ skill_log   │ ── learnings back to RAG
                │ ≥ thresh │  │ (situation, │
                └────┬─────┘  │  tactic,    │
                     │        │  outcome)   │
              autonomous?     └─────────────┘
              ┌──┴──┐
              ▼     ▼
        execute   ┌──────────────┐
                  │ Telegram bot │── human signoff ──▶ execute
                  └──────────────┘

Architecture (TL;DR)

• Discovery Council (3 new + 4 shared roles, no CEO) — debates what to build using the 3-arbitrages rubric (Capacity / Information / Distribution).
• Execution Council (8 roles) — debates how to build a chosen opportunity through an 11-stage protocol with adversarial double + Skeptic veto.
• KPI Gate State Machine — Postgres-backed approval queue with Telegram signoff for non-autonomous actions.
• Skill Library — Voyager-inspired, pgvector-backed (situation, tactic, outcome) store the councils RAG over.

Detailed specs in docs/:

• prior-art.md — what we reuse vs build
• success-factors.md — 10 structural moves to raise success odds
• human-in-the-loop.md — your role, time budget, taste anchors, conviction log
• business-frameworks.md — BMC + Lean Canvas + VPC + JTBD + 22 Laws + 9 more
• discovery-council.md — opportunity sourcing & scoring
• architecture.md — Council Protocol, Gates, Skill Library, build order

Project rules in CLAUDE.md (read by Claude Code in every session). The hardest rule: research before build — see ~/.claude/commands/prior-art.md and docs/prior-art.md.

Quick start

Prereqs: Node ≥20, pnpm ≥9, Docker (for Postgres+pgvector). LiteLLM proxy at http://localhost:4000 (per global rule ~/.claude/rules/common/llm-routing.md).

# Install
pnpm install

# Start Postgres+pgvector
pnpm db:up

# Run migrations (after Postgres is up)
pnpm db:migrate

# Typecheck & tests
pnpm validate

# Smoke test (verifies LiteLLM connectivity if available, dry-run otherwise)
pnpm smoke

# Run the Council demo on a synthetic objective (no money touched)
pnpm council:demo

Environment

Copy .env.example to .env for local dev with placeholder values. For real use, resolve .env.tpl with the 1Password CLI:

op inject -i .env.tpl -o .env.resolved

Per ~/.claude/rules/common/security.md, never commit .env or .env.resolved.

Repo layout

src/
├── llm/         @thio/llm-client wrappers + alias constants
├── council/     Execution Council (11-stage protocol, 8 roles)
│   ├── stages/  intake, diverge, evidence-gate, steel-man, propose, adversarial-double, decide, gate-check, execute, verify, learn
│   └── roles/   ceo, cmo, engineer, cfo, skeptic, compliance, anthropologist, brand-steward
├── discovery/   Discovery Council (3-arbitrages scoring, opportunity surface)
│   ├── roles/   scout, capability-mapper, distribution-analyst
│   ├── scoring/ arbitrages, fit, gates
│   └── sources/ reddit, hackernews, indiehackers, producthunt, github-trending, jobs, newsletters
├── gates/       KPI Gate State Machine + Telegram approval bot
├── skills/      Skill Library (read/write + pgvector RAG)
├── tools/       External integrations (Postiz, content pipeline, analytics, compliance)
├── infra/       LangGraph/Inngest/Langfuse wiring
└── common/      shared types, errors, logger

Test of success

Hours of human time per dollar of revenue, plotted weekly. Trending down → working. Flat → product-market-fit problem. Trending up → architecture leak.

See docs/human-in-the-loop.md § "The success metric".