Features
Everything you need from an AI agent. Nothing you don't want from a privacy tool.
Encrypted at Rest
Every piece of data Aivyx stores is encrypted with ChaCha20Poly1305 authenticated encryption. Keys are derived using HKDF-SHA256 with a unique salt per purpose — compromising one key doesn't expose data encrypted with another.
- Master passphrase → Argon2id → HKDF per-purpose keys
- Separate keys for memory, audit, sessions, secrets
- Key material zeroed on process exit via
secrecycrate
Persistent Memory
Aivyx builds a semantic knowledge graph from your conversations. It extracts facts as subject-predicate-object triples and stores them encrypted. Your agent genuinely learns and remembers over time.
- Automatic fact extraction from conversations
- Semantic triple store with encrypted storage
- User profile learning (preferences, timezone, tools)
- Full control:
aivyx memory list, search, delete
Multi-Provider LLM
Use any LLM provider — from fully local to cloud. Switch between them per-task. Configure different providers for different purposes.
Tamper-Proof Audit
Every action your agent takes is recorded in an HMAC chain. Each entry signs the previous hash, creating a tamper-evident log. If anyone modifies a single entry, the chain breaks.
$ aivyx audit verify
✓ Chain integrity verified
✓ 1,247 entries validated
✓ No tampering detected 22 Built-in Skills
Structured workflows that produce professional-quality results. From code review to financial analysis, each skill follows a proven methodology.
46 Built-in Tools + MCP
A comprehensive toolkit spanning 4 phases of development. Plus the Model Context Protocol for infinite extensibility.
Core (22 tools)
File read/write/delete/move/copy, shell, web search, HTTP fetch, grep, glob, directory list, project tree/outline, text diff, git (status/diff/log/commit), JSON parse, hash, datetime, env
Analysis (Phase 11A)
CSV query, math eval, config parse, regex replace, text statistics, sentiment analysis, entity extraction, PII detection, risk matrix, image metadata
Documents (Phase 11B)
Document extraction, chart generation, diagram authoring, template rendering, markdown export, HTML-to-markdown
Infrastructure (Phase 11D)
Log analysis, compliance checking, file patching, archive management
$ aivyx mcp list --command "npx -y @mcp/server-github"
Tools discovered:
• create_issue (owner, repo, title, body)
• search_repositories (query)
• get_file_contents (owner, repo, path)
• create_pull_request (owner, repo, title, head, base) The Nonagon Team
Nine specialist agents collaborate on complex missions. Each has its own persona, tools, skills, and capability boundaries. The Coordinator decomposes goals, delegates to the right specialist, and synthesizes results.
Federation
Agents on different instances can communicate and share memory. Send messages to federated peers, search memories across instances, and build truly distributed AI workflows.
- Cross-instance agent messaging with Ed25519 signature verification
- Federated memory search across instances
- Trust policies with allowed scopes and tier limits per peer
- Multi-region failover with capability-aware peer selection
DAG Task Execution
Move beyond sequential planning with a full directed acyclic graph execution engine. Steps run in parallel when independent, with results forwarded to dependents.
- Topological sort with wavefront parallel execution
- Reflection loops — LLM-as-judge with automatic step re-insertion
- Human-in-the-loop approval checkpoints with timeout escalation
- Dynamic agent spawning mid-session with auto-cleanup
Step 1: Research ────┐
Step 2: Analyze ──────┤→ Step 4: Synthesize → Step 5: Review
Step 3: Code (needs 1)┘ (needs 1,2,3) (approval) Voice & Vision
Talk to your agent naturally with real-time WebSocket voice. Analyze images and documents with vision models. All modalities feed into encrypted multimodal memory.
- WebSocket voice: Listening → Processing → Speaking state machine
- STT via Whisper or Ollama, TTS via OpenAI or edge-tts
- Barge-in interruption with CancellationToken
- Vision across Claude, OpenAI, and Ollama providers
- PDF, XLSX, CSV document extraction pipeline
- Multimodal memory with image attachment embeddings
A2A + MCP Protocols
First-class support for both Google A2A (agent-to-agent) and MCP (agent-to-tool). Backed by 50+ companies, these are the emerging industry standards.
- A2A Agent Card, JSON-RPC task lifecycle, SSE streaming, push notifications
- MCP with OAuth 2.1 + PKCE, sampling, elicitation, hot-reload
- Server registry integration (Smithery.ai, mcp.run)
- Federation for cross-instance agent collaboration
Multi-Tenancy & Governance
Ship to production with confidence. Full multi-tenant isolation, enterprise authentication, and cost governance — all encrypted by default.
- Per-tenant HKDF key derivation with isolated directory trees
- RBAC with 4 roles: Billing, Viewer, Operator, Admin
- OIDC SSO with group-to-role mapping
- Cost ledger with per-agent/tenant budgets and daily/monthly limits
- Model routing by purpose (planning, execution, embedding)
- Webhook triggers with HMAC-SHA256, multi-stage workflow engine
- Kubernetes Helm chart with HPA, Ingress, PVC, secrets
Agents That Learn
Aivyx agents improve over time. Outcome tracking, feedback loops, and knowledge graph evolution create a system that gets smarter with use.
- GraphRAG — knowledge graph with BFS, community detection, entity search
- Agentic RAG — retrieval router with quality evaluation and multi-source synthesis
- Memory consolidation — clustering, merge, decay pruning of stale memories
- Outcome tracking with per-tool and per-role success rates
- Planner feedback loops identifying successful tool combinations
- Specialist recommendation learning with historical weight bonuses
Hardened & Observable
Battle-tested with 1,576 tests across both repos. OWASP-aligned security, structured telemetry, and CI/CD-ready infrastructure.
- Prompt injection defense — 3-layer sanitization (ChatML, tool output, webhooks)
- Tool abuse detection — sliding-window anomaly on frequency and scope escalation
- Capability audit reports flagging overly permissive grants
- OpenTelemetry + Prometheus + W3C Trace Context propagation
- OpenAPI 3.1.0 spec covering all 90+ endpoints
- Chaos testing (fault injection middleware) + k6 load tests
- Horizontal scaling with session affinity strategies