CIRIS replaces apps like ChatGPT and Grok everywhere you need AI. It's the same chat you already expect — but it shows its reasoning, escalates to a human when it's unsure, keeps your data private, works in 29 languages, and runs on your own device. Open source, free, no ads, no growth-at-all-costs pressure.

Desktop & self-host: pip install ciris-agent — see Run it yourself. Sign in with Google for the free hosted CIRIS model, or bring your own key (OpenAI, Anthropic, Groq, Together.ai, or a local model).

Why CIRIS

• It shows its work. Every answer passes ethical, common-sense, domain, and reasoning-fragility checks — and you can see why it said yes or no, not just the answer.
• It defers to you. When a decision is uncertain, CIRIS escalates to a designated human ("Wise Authority") instead of guessing.
• Private by design. Runs on your device. The hosted CIRIS proxy stores nothing — your prompts are not logged and never train a model.
• Speaks your language. The entire ethical-reasoning system — not just the buttons — operates in 29 languages.
• Auditable and open. AGPL-3.0, cryptographically signed decisions, a tamper-evident audit trail, and a public ethical framework anyone can review: the Constitution (the Accord folded into it; PDF).

Not a replacement for humans — a tool that knows its limits.

How it works

CIRIS wraps every AI response in a reasoning pipeline: multiple evaluation passes for ethics, common sense, domain knowledge, and reasoning fragility (it flags answers that lean on a single weak source). Uncertain calls defer to designated humans. Every decision is written to a hash-chained audit trail. Today CIRIS powers Discord community moderation in production at agents.ciris.ai; the architecture is built to scale to settings like education and healthcare.

The design is described in the Coherence Ratchet paper.

Run it yourself

pip install ciris-agent
ciris-agent                       # desktop app + local API server
ciris-agent --adapter discord     # or run it as a Discord bot

Server install (agent + web UI) — download, optionally inspect, then run:

curl -fsSLO https://ciris.ai/install.sh
# (optional) read install.sh to see what it does, then:
bash install.sh

Security note: piping curl ... | bash runs unreviewed code straight from the network. Downloading first lets you inspect the script before executing it.

For developers

Under the consumer app, CIRIS is a type-safe, auditable AI agent framework — 22 core services on a 6-bus message architecture, 200+ API endpoints, 4 GB RAM target, 10,000+ tests. Extend it with adapters, run it headless, or embed it.

• Documentation Hub — everything, organized
• Architecture · API Reference · Quick Start
• Contributing · Security · For AI assistants

Honest read

CIRIS is real and running in production, but young — version 2.x, under active development. It proves an AI is accountable, not that it is correct: the reasoning is made visible so you can judge it yourself. It does not give medical advice or substitute for professional care. It is not magic — it is an ordinary chat assistant, plus the accountability machinery that closed apps don't give you.

CIRIS — Core Identity, Integrity, Resilience, Incompleteness, and Signalling Gratitude Copyright © 2025 Eric Moore and CIRIS L3C · AGPL-3.0 · Release Notes · Issues