Missy¶

Security-first, self-hosted AI assistant for Linux. Production-grade agent platform with strict policy enforcement, encrypted vault, multi-provider support, voice channel, and full auditability.

Get Started View on GitHub

Security First¶

Every capability is disabled by default. Network, filesystem, shell, and plugin access require explicit opt-in. Three-layer policy engine enforces rules on every request. ChaCha20-Poly1305 encrypted vault for secrets. Container-per-session sandbox for OS-level isolation, Ed25519 agent identity for signed audit trails, and trust scoring (0--1000) for provider and tool reliability tracking.

Multi-Provider¶

Switch between Anthropic, OpenAI, and Ollama (local models) — even at runtime. API key rotation, model tiers (fast/primary/premium), and automatic fallback when a provider is down.

Voice Native¶

WebSocket voice channel with dedicated Raspberry Pi edge nodes. Wake word detection, local STT (faster-whisper), local TTS (Piper). Per-node policy modes with PBKDF2 device authentication.

Agentic Runtime¶

Multi-step tool loop with circuit breaker, checkpoint/recovery, cost tracking, and budget caps. Sub-agents, learnings extraction, and self-tuning prompt patches. AI Playbook auto-captures successful tool patterns and promotes them to skills. Sleep Mode consolidates context when the token window fills. Attention System tracks urgency, focus, and topic continuity across turns. Interactive approval TUI surfaces policy-denied operations for real-time operator approval with session-scoped "allow always" memory.

Full Auditability¶

Every action logged as structured JSONL — network requests, file access, shell commands, tool calls. OpenTelemetry export for production observability. No silent failures.

Extensible¶

Built-in tools, skills, plugins, and MCP server integration. Digest-pinned MCP connections for supply chain safety. Config presets, auto-migration, plan/rollback for operations. FAISS vector memory for semantic search across conversation history. SKILL.md dynamic discovery for cross-agent skill portability. Async Message Bus with topic wildcards and priority queuing for event-driven subsystem coordination.

Quick Install¶

curl -fsSL https://raw.githubusercontent.com/MissyLabs/missy/master/install.sh | bash

Or install manually:

pipWith voice supportWith OpenTelemetry

git clone https://github.com/MissyLabs/missy.git
cd missy
pip install -e .
missy setup

pip install -e ".[voice]"

pip install -e ".[otel]"

Architecture¶

CLI / Discord / Webhook / Voice
         │
    AgentRuntime
    ├── InputSanitizer + SecretsDetector + PromptDriftDetector
    ├── PolicyEngine (network / filesystem / shell / REST L7)
    ├── AgentIdentity (Ed25519) + TrustScorer (0-1000)
    ├── CircuitBreaker + RateLimiter
    ├── ContextManager (token budget) + MemoryConsolidator (sleep mode)
    ├── AttentionSystem (alerting / orienting / sustained / selective / executive)
    ├── ProviderRegistry (Anthropic / OpenAI / Ollama)
    ├── ToolRegistry + MCP Manager + SKILL.md Discovery
    ├── Memory + Learnings + Vector Memory (FAISS) + MemorySynthesizer
    ├── Playbook (auto-capture → skill promotion)
    ├── MessageBus (async events, topic wildcards, priority queue)
    ├── ApprovalGate + InteractiveApproval TUI
    ├── ContainerSandbox (Docker isolation)
    └── AuditLogger + OpenTelemetry

Channels¶

Channel	Description
CLI	Interactive REPL and single-shot `missy ask`
Discord	Full Gateway API with DM/guild policies, slash commands
Webhook	HTTP ingress for automation pipelines
Voice	WebSocket server + Raspberry Pi edge nodes with wake word

MissyLabs Missy Edge