Skip to content

Roadmap

Phase 0: Weekend MVP (Complete)

  • Single-machine AI assistant with tool calling
  • ReAct agent loop
  • Hardware auto-detection
  • Interactive CLI and REST API

Phase 1: Multi-Machine Orchestration (Complete)

  • Phase 1.1 (Complete): Cluster foundation
  • Cluster configuration schema
  • Remote LLM client
  • Health monitoring and node selection

  • CLI commands for cluster management

  • Phase 1.2 (Complete): Discovery & Health

  • mDNS auto-discovery for local networks
  • Periodic health monitoring
  • Circuit breaker pattern

  • Retry logic with exponential backoff

  • Phase 1.3 (Complete): Smart Routing

  • Task complexity classification
  • Context-aware routing decisions
  • Automatic tier selection

  • Integration with agent loop

  • Phase 1.4 (Complete): Polish & Monitoring

  • Dynamic node management (add/remove nodes at runtime)
  • Performance metrics collection and tracking
  • REST API metrics endpoint
  • CLI metrics command

Phase 2: Memory & Context (Complete)

  • Phase 2.1 (Complete): Conversation Memory
  • SQLite-based conversation persistence
  • Session management and lifecycle
  • Token-based context windowing
  • Multi-turn conversations with history recall

  • Optional memory (backward compatible)

  • Phase 2.2 (Complete): Semantic Search & RAG

  • Vector embeddings with sentence-transformers (privacy-first, local)
  • ChromaDB vector store for similarity search
  • Semantic search across conversation history
  • RAG (Retrieval-Augmented Generation) for context-aware responses
  • Cross-session knowledge retrieval

Phase 3: Voice & Multi-Modal (Complete)

  • Whisper STT integration (speech-to-text)
  • TTS integration (Piper fast, Coqui high-quality)
  • Voice client (push-to-talk)
  • Voice API endpoints (REST + WebSocket)
  • Real-time audio processing
  • Cross-platform audio I/O (macOS, Linux, Windows)

Phase 4: Security Layer (Complete)

Phase 4.1-4.4 (Complete): Core security infrastructure

  • MCP Protocol Base - JSON-RPC 2.0 protocol implementation
  • Docker Container Manager - Container lifecycle management with resource limits
  • MCP Gateway Server - Centralized gateway for tool execution routing
  • Audit Logging System - SQLite-based comprehensive audit trail
  • Event tracking (requests, responses, errors)
  • Tool call logging with parameters and results
  • Security decision logging
  • Sensitive data redaction (API keys, passwords, tokens)
  • Secret Management - Multi-backend credential vault
  • HashiCorp Vault integration (production)
  • SOPS file encryption (small teams)
  • Environment variables (development)
  • Secret scanning and detection
  • Automatic secret injection into containers
  • Network Isolation - Per-container egress filtering
  • Docker network isolation
  • iptables-based egress rules
  • Domain allowlists
  • DNS query filtering
  • Connection attempt logging

Phase 4.5 (Complete): Human-in-the-Loop Gates

  • Risk-based classification (LOW/MEDIUM/HIGH/CRITICAL)
  • Centralized approval management with timeout handling
  • CLI approval prompts with rich console formatting
  • API approval support with web-compatible data structures
  • Gateway integration with HITL checks before tool execution
  • Audit integration for all approval decisions
  • Default-deny safety on timeout

Phase 4.6 (Complete): Browser Container with Pre-Authentication

  • Playwright-based browser automation with session isolation
  • Pre-authentication flow with credentials injected from vault backend
  • Accessibility-based interaction (semantic tree instead of raw HTML/DOM)
  • Six browser tools: navigate, click, type, read, screenshot, close_session
  • Password field protection (auto-deny typing into password/secret fields)
  • HITL integration with 16 risk classification rules
  • Session management with timeout and action-count expiration

Phase 4.7 (Complete): Code Execution Sandbox

  • gVisor-based sandboxed code execution (Python, JavaScript, shell)
  • Six MCP tools: execute, install package, write/read/list files, destroy sandbox
  • HITL risk classification rules (network + dangerous patterns = CRITICAL)
  • Gateway routing for all sandbox tools
  • Docker container with FastAPI MCP server

Phase 4.8 (Complete): End-to-End Security Integration

  • All security components wired through MCP Gateway
  • Per-tool container isolation with resource limits
  • Audit logging across all tool invocations
  • HITL gates integrated at gateway level

Phase 5: Intelligence & Privacy (Complete)

  • ML-based anomaly detection with Isolation Forest (per-agent models)
  • Threat scoring, threat intelligence integration
  • Trust manager with historical risk scoring and auto-approval
  • Secret rotation with zero-downtime and emergency rotation
  • Certificate pinning, deep packet inspection, protocol filtering
  • SIEM integration (Splunk, Elasticsearch, Datadog)
  • Alert rules engine, compliance reports, security dashboard
  • Privacy Router: hybrid local/cloud AI with PII detection
  • Multi-model collaboration patterns

Phase 6: Advanced Security (Experimental)

Note: Phase 6 features are implemented but experimental. They have not been validated in production environments and may require additional hardening before enterprise deployment.

  • Hardware security module integration — Software simulation only; requires specific hardware (TPM/SGX/SEV-SNP) for production use
  • Enhanced isolation mechanisms — Designed but not validated at scale
  • Zero-knowledge proof support — Integrated with audit pipeline for privacy-preserving compliance proofs
  • Distributed cryptography primitives — Shamir secret sharing and MPC implemented; requires security audit

Code Quality

  • Mypy strict mode (Complete): Zero errors across 155 source files with strict = true, enforced as a blocking check in CI and pre-commit

Documentation