Assistant for Realtime Ethical Thought and Evaluation

An ethics-first, transparent reasoning assistant — built to be self-hosted by anyone.

ARETE is an AI assistant that tries to show its work.

Most assistants today give you polished answers but hide how they got there. ARETE does the opposite: it focuses on clarity over speed or persuasion. For each response, it aims to surface:

• how confident it is,
• what sources it relied on,
• what trade-offs it considered,
• and what ethical constraints were in play.

The goal is not to replace human judgment, but to make it easier for people and communities to inspect, question, and correct the system.

• Ethics-first design – Every feature should help people think more clearly about what matters.
• Transparency & provenance – Reasoning and sources should be traceable, not hidden in a black box.
• Humility & pluralism – The assistant should express uncertainty and make room for multiple perspectives.
• Auditability – Decisions and responses should be loggable and explainable.
• Responsiveness – Transparency should invite discussion and correction.
• Human oversight – ARETE can guide reflection, but it should never be treated as an oracle.
• Open & self-hostable – Anyone should be able to inspect, modify, and run their own instance.

Explore the public site for more details and a quick demo!

ARETE is made up of three small services that work together:

• Discord bot
  The conversational interface. It listens in Discord, talks directly to the AI model to generate replies, and sends trace metadata to the backend for safekeeping.

• Web interface
  The public-facing site and explanation viewer. It displays documentation and trace reports, and forwards API requests to the backend in a controlled way.

• Backend API
  The system’s memory and guardrail layer. It stores response traces, serves runtime configuration, verifies CAPTCHA challenges, enforces rate limits, and exposes audit data.

In production, these services are deployed separately but are designed to behave the same way locally and in the cloud.

Locally, the web service proxies /api/* to the backend. On Fly, it reaches the backend over the internal Fly network.

git clone https://github.com/arete-org/arete.git
cd arete

pnpm install

If pnpm isn't available yet, run corepack enable once (Node 16.10+), then re-run pnpm install.

cp .env.example .env

At minimum:

DISCORD_TOKEN=...
DISCORD_CLIENT_ID=...
DISCORD_GUILD_ID=...
DISCORD_USER_ID=...
OPENAI_API_KEY=...
TRACE_API_TOKEN=...
INCIDENT_PSEUDONYMIZATION_SECRET=...

See .env.example for the full list and descriptions of optional settings.

Start the backend and web interface:

pnpm start:dev

In another terminal, start the Discord bot:

pnpm start:bot

• pnpm start:dev - Backend + web (no bot).
• pnpm start:bot - Discord bot only (requires bot env vars).
• pnpm start:backend - Backend only (no web or bot).

• pnpm start:prod-test - Build backend and run production server.
• pnpm dev:prod-test - Production backend + web dev server (no bot).

• pnpm build - Build all workspace packages.
• pnpm pre-review - ARETE tags + type-check + lint.

• Cloudflare Turnstile (abuse prevention)
  Turnstile protects public endpoints from abuse.
  If both keys are set, CAPTCHA is enforced.
  If neither key is set, CAPTCHA is skipped.

TURNSTILE_SITE_KEY=...
TURNSTILE_SECRET_KEY=...

• Cloudinary (image uploads)
  If Cloudinary credentials are provided, images can be uploaded and referenced in traces.
  If not, the system falls back to attaching images directly in Discord.

CLOUDINARY_CLOUD_NAME=...
CLOUDINARY_API_KEY=...
CLOUDINARY_API_SECRET=...

Response traces are stored in SQLite:

PROVENANCE_SQLITE_PATH=/data/provenance.db

On Fly.io, /data is backed by a persistent volume. On other hosts, point this path at a durable directory.

The repository supports multi-service deployment:

• Docker Compose (local)
• Fly.io (three separate apps: backend, web, discord-bot)

Deployment configuration and scripts live under:

deploy/

See deploy/README.md for details.

ARETE is dual-licensed under:

• MIT
• Hippocratic License v3 (HL3-CORE)

See docs/LICENSE_STRATEGY.md for details.

Governance and contribution guidelines are still being drafted.

For now, thoughtful discussion, critique, and experimentation are welcome via issues and discussions.