Auxiora
Security-first personal AI assistant platform — open source, cross-platform, model-agnostic, with encrypted credential vault and plugin marketplace.
Problem
Personal AI assistants have exploded in popularity but ship with catastrophic security defaults — plaintext credentials, no authentication, secrets in the model's context window. The most popular one had 17,903 instances exposed on Shodan with API keys accessible to anyone.
Approach
Built a modular TypeScript monorepo from scratch with security as the foundation. Encrypted credential vault (AES-256-GCM + Argon2id), architectural separation so secrets never touch the AI model, zero-trust user pairing, and tamper-evident audit logging. Multi-platform messaging adapters, model-agnostic provider system with failover, plugin marketplace with sandboxed execution, and daemon mode for always-on operation.
Result
60+ packages, 1,700+ tests passing. Runs across macOS, Linux, and Windows 11. Supports Discord, Telegram, Slack, SMS/WhatsApp, and WebChat. Model-agnostic across Claude, GPT, Gemini, and local LLMs. Plugin marketplace live with community extensions.
Stack
What I'd do differently
Would build the Rust native performance layer (napi-rs) from day one instead of planning it for v2.0. The TypeScript crypto implementations work but native addons would give constant-time operations and eliminate OpenSSL dependency.
Auxiora is a personal AI assistant platform built for people who want the productivity gains of a 24/7 AI assistant without becoming a security liability. It runs on your devices, connects to the messaging platforms you already use, and encrypts everything.
Architecture
The system is a modular monorepo with clear separation of concerns:
- Gateway — WebSocket + HTTP server with rate limiting, input sanitization, and JWT authentication
- Agent Runtime — Provider factory with failover, personality engine, and tool system
- Channel Adapters — Standardized interface for Discord, Telegram, Slack, SMS/WhatsApp, and WebChat
- Security Layer — Encrypted vault, tamper-evident audit logs, zero-trust pairing, prompt injection detection
- Plugin System — Sandboxed execution, marketplace, webhook ingestion, and cron-based scheduled tasks
Key differentiators
- Secrets never touch the model — The AI requests actions, a separate execution layer injects credentials. Prompt injection cannot extract API keys.
- Personality system — SOUL.md, USER.md, and AGENTS.md define how the assistant thinks, communicates, and what it is allowed to do. It learns your patterns over time.
- Model agnostic — Primary + fallback provider configuration. Swap between Claude, GPT, Gemini, or local LLMs without changing workflows.
- Platform extensibility — Write a new channel adapter in an afternoon. Plugin marketplace for community workflows. MIT-licensed.
Learn more at auxiora.ai.
Interested in working together?
Get in touch