Konsey: a multi-LLM council where a model can't verify its own output

a single LLM to grade its own answer is a conflict of interest. So I built Konsey — a small, MIT-licensed orchestration CLI that runs several providers as a council.

How it works: independent providers each plan, an adversarial critic attacks the plans, and a *different* provider verifies the result. The rule is hard: the producer never verifies its own output. It's a 9-state loop (plan → critique → synthesize → execute → verify → decide) with an append-only DuckDB audit trail.

Why it might matter to you:

- Vendor-neutral: Claude, Codex, Gemini — or a local Ollama / llama.cpp model. Config-driven, one line to swap.
- Local-first: run fully offline; nothing leaves your machine (good for PHI/regulated work).
- Layered, opt-in safety: strict/medium/weak levels, secret + PHI scanning.
- Chat connectors (Telegram, Notion, Slack, WhatsApp) with the same risk gate on every channel.

Install (two deps, langgraph + duckdb):

curl -fsSL https://raw.githubusercontent.com/eMediquality/konsey/v0.2.0/install.sh | KONSEY_REF=v0.2.0 bash

Repo + docs: https://github.com/eMediquality/konsey — feedback and contributors very welcome.