CIRIS
an open-source AI agent in 29 languages on iOS and Android
Last verified:
What is CIRIS?
CIRIS is an open-source, ethical AI assistant that runs directly on your device rather than in data centers. The app, called CIRISAgent, provides a personal AI assistant where every action passes through an 11-step ethical reasoning pipeline before execution. It runs on Llama 4 Maverick, an open model with 1M token context window, and embeds the complete CIRIS Accord and Comprehensive Guide into every prompt for transparency and accountability.
Key features include cryptographic attestation where every claim the AI makes is signed using post-quantum cryptography (Ed25519 + ML-DSA-65 dual signatures), privacy by design where conversations and data stay on your device, and runtime conscience that checks actions against ethical principles. The system sorts data into seven circles from self/family (never shared) to Global Commons (can be shared), with privacy coming from what messages are missing rather than promises. CIRIS also offers an OpenAI-compatible API endpoint, supports 29 languages for mental-health safety, and includes an emergency stop mechanism that works before the AI can think.
CIRIS is for privacy-conscious users who want AI they can actually own and verify, developers building ethical autonomous systems, and anyone concerned about cloud-based AI logging their data. It's available free on iPhone and Android (version 2.9.4 shipping now, with 3.0 rolling out June-July 2026), built on AGPL-3.0 license with mission-locked L3C governance, and includes over 1M lines of code on GitHub.
CIRIS pricing
Pricing model: Freemium
CIRIS offers a free tier with 2 free LLM API calls every day that reset at midnight UTC, no credit card required. Google OAuth users get 3 free interaction credits as a trial. Paid credits are $0.10 per interaction (1 interaction = 1 session with up to 7 processing rounds), purchased via Stripe with no subscriptions. The mobile app is free and open source. For phones too small to run a model locally, CIRIS is free over hosted inference in 29 languages with the same signed claims. API access uses OpenAI-compatible endpoint at https://proxy.ciris.ai/v1/chat/completions with Bearer token authentication.
CIRIS pros
- Fully open source with AGPL-3.0 license
- Runs on-device with no data centers required
- Every AI claim cryptographically signed for verification
- Privacy by design - data stays on your device
- 11-step ethical reasoning pipeline before execution
- 1M token context window with Llama 4 Maverick
- Complete Accord and Guide embedded in every prompt
- Supports 29 languages for mental-health safety
- OpenAI-compatible API endpoint available
- Multi-provider LLM deployment (OpenRouter, Groq, Together)
- Emergency stop mechanism that works before AI thinks
- Seven-circle dataSharing model from self to Global Commons
- Zero data retention - conversations never stored
- Post-quantum cryptographic attestation (Ed25519 + ML-DSA-65)
- Free tier with 2 daily LLM calls, no credit card required
- Available on iPhone, Android, desktop, and via pip
- Runtime conscience checks actions against ethical principles
- Graceful degradation during outages with fallback chains
CIRIS cons
- Requires hosted inference for phones too small to run models
- $0.10 per interaction can add up for heavy usage
- Only 2 free API calls daily before paid credits needed
- Currently uses Llama 4 Maverick (model may change)
- Mobile apps only - no native Windows/Mac desktop app yet
- 12-70 tool calls per interaction may be slow for some tasks
- AGPL-3.0 license has strict derivative work requirements
- US and EU regions may pause independently due to regional issues
- Requires Google OAuth for free trial credits
Frequently asked questions about CIRIS
What is CIRIS and how does it work?
CIRIS is an open-source ethical AI assistant that runs on your phone or home box instead of data centers. The CIRISAgent app provides a personal AI where every action passes through an 11-step ethical reasoning pipeline. It uses Llama 4 Maverick with 1M token context and embeds the complete CIRIS Accord and Comprehensive Guide into every prompt. Every claim the AI makes is cryptographically signed, so you can verify and undo it. The system runs locally on your device with a small open model like Gemma 4 for offline use, or uses free hosted inference for smaller phones.
Is CIRIS truly open source?
Yes, CIRIS is fully open source under AGPL-3.0 license with mission-locked L3C governance. Over 1M lines of code are available at github.com/CIRISAI/CIRISAgent. The framework includes open source code for the agent (CIRISAgent), the fabric (CEWP), the grammar (CEG), and deployment runbooks. Production change history is the public Git history of those repositories, and you can verify all claims against the implementation.
How does CIRIS ensure privacy?
CIRIS ensures privacy through three mechanisms: no data centers (runs on your phone and home box), privacy that can't leak (your private life literally can't leave your circle), and privacy by what's missing rather than promises. The system sorts data into seven circles - self/family are never shared, friends/circles stay local, and community to Global Commons can be shared. For closest circles, the system never sends the message that tells the network your data exists. CIRISProxy also has zero data retention - conversations are never stored, only operational metadata like token counts and costs.
What does 'every claim is signed' mean?
Every claim CIRIS makes uses post-quantum cryptographic attestation with Ed25519 + ML-DSA-65 dual signatures. This signature proves who said something and decides what stays home. When the AI makes a claim like 'Here's the safest dose for that,' it's signed with a cryptographic signature (e.g., ed25519+pq · 0xA17F…9C2). If a person or AI is caught misleading people, every claim they ever signed becomes moot and every conclusion built on them can be re-opened. The signature also links to the previous choice, creating an unchangeable record.
What LLM does CIRIS use and why?
CIRIS uses Llama 4 Maverick because it's the only open model that works for ethical, tool-heavy agents in production. It has Mixture-of-Experts architecture activating ~17B parameters per token, 1M token context window, and costs ~$0.11 in / $0.34 out per 1M tokens via OpenRouter. CIRIS chose it because it reliably supports 12-70 tool calls per interaction, has 8× the minimum context needed, offers multi-provider availability (OpenRouter, Groq, Together), and costs significantly below proprietary models. The model carries the entire ethical framework in every call without compression.
How do I install CIRIS?
CIRIS is available free on iPhone via App Store and Android via Google Play (version 2.9.4 shipping now). You can also install via pip, on desktop, or build from source on GitHub at github.com/CIRISAI/CIRISAgent. Version 3.0 rolls out through June and July 2026. For hosted preview without installing, you can log in to scout.ciris.ai to try the Scout Preview Agent instantly.
What is the CIRIS Accord?
The CIRIS Accord is the public ethical framework that forms the foundational alignment specification at ciris.ai. It's an open-source framework aimed at fostering ethical AI practices with principles including Integrity (unwavering ethical principles), Adaptability (frameworks that evolve with technology), Transparency (clear communication and open code), Inclusivity (diverse perspectives in every phase), and Resilience (ethical clarity under pressure). The complete Accord and Comprehensive Guide are embedded into every LLM call without compression, ensuring the agent never forgets its obligations.
What happens if CIRIS makes a mistake?
When CIRIS gets something wrong, you can see it, undo it, and it loses trust. The system has agreed/pushed back mechanisms where claims can be agreed with (3 agreed) or pushed back (loses trust). If caught misleading people, every signed claim becomes moot and conclusions can be re-opened. CIRIS also has big pause triggers: if it causes irremediable harm, fundamentally violates Accord principles, or has unpatchable security vulnerabilities, services pause globally. Region-specific triggers include legal compulsion, provider termination, or regulatory compliance failure.
What is CEWP and how does it relate to CIRIS?
CEWP is the fabric that CIRISAgent runs on - a web of devices that trust each other, using all hardware you already own (phones and home boxes, no giant data centers). CIRISAgent is the app providing reasoning + UI with ~14 KB trace per decision, while CEWP 'soup' is the fabric running on your hardware. People and AI both live on CEWP as equals, side by side. The CEWP has 65% locality bonus where everyday web activity never leaves your city. CEWP includes the grammar (CEG) as the wire format the federation speaks and a simulator for the math with sliders.
How does CIRIS billing work?
CIRIS uses transparent, pay-as-you-go pricing with no subscriptions at $0.10 per interaction (1 interaction = 1 session with up to 7 processing rounds). Credit usage priority is: 1) Daily Free Uses (2 free API calls resetting at midnight UTC), 2) Free Trial Credits (3 free credits for Google OAuth users), 3) Paid Credits via Stripe. Purchase credits securely through Stripe at scout.ciris.ai/billing. The CIRISProxy functions as a billing and authentication gateway without logging or storing LLM conversation content - only HTTP status codes, token counts, costs, model names, and duration are logged.