Multi-model, event-driven personal OS with proactive intelligence, layered security, and self-improving behavior.
Click any section for a plain-English explanation.
Command Interfaces
You can talk to Miss Chief through messaging apps or AI coding tools. Every message, regardless of where it comes from, flows through the same intelligence and safety pipeline. The system routes your request to the right AI model and responds in the same channel you used.
Defense in Depth
Before the AI does anything, every action passes through 9 layers of security. A killswitch enforces hard limits on message rates and allowed hours. An injection scanner detects manipulation attempts. An output sanitizer strips secrets. Different security profiles apply depending on who the AI is talking to. Chat isolation prevents data leaking between conversations. Actions are classified into 4 approval tiers, from silent reads to explicit human approval.
Event-Driven Lifecycle
12+ hooks across the full lifecycle. Self-correcting: captures patterns, corrections, and context automatically at every stage.
The system hooks into every phase of a conversation. When a session starts, it loads memory and personality. Before each message, it checks for triggers and loads relevant data. After tools run, it tracks changes. Before context compression, it snapshots critical data. When the session ends, it extracts a structured summary for future recall.
Core Systems
Session memory with semantic + temporal search, knowledge graph for entities and relations, semantic memory for learnings and habits.
Two-tier heartbeat: silent scan, narrative only on critical. Multi-agent monitoring with event-driven triggers.
Auto-extracts NEVER/ALWAYS rules. Model-agnostic across all providers. Nightly eval, fix, verify, keep or rollback.
The system connects to 15+ services to stay informed. It reads and writes WhatsApp, Gmail, Slack. It syncs calendars, pulls meeting transcripts, researches across YouTube, Reddit, and web search. All data synced in parallel before every briefing, so the AI reasons on fresh information.
Three memory systems working together. Session memory remembers every conversation, searchable at 50:1 compression. A knowledge graph tracks people and projects. Semantic memory stores behavioral patterns. The corrections engine learns from mistakes and proposes permanent system changes when the same issue recurs.
Skills are modular capabilities the AI can use. Morning and evening skills run daily briefings automatically. Content skills manage a publishing pipeline. Research skills dispatch multiple agents in parallel. Skills can be invoked manually, triggered by automation, or activated proactively when the AI detects they are needed.
Autonomous Systems
Silent by default. Zero LLM on normal runs. Haiku-narrated alerts only when something genuinely matters.
Cost-aware routing. Output enforcement strips AI patterns before delivery. WhatsApp, Slack, Telegram, Email, Drive.
External model as judge. Budget-capped at 5 iterations. Rollback-safe.
Three watchdogs running in the background. One monitors operations (urgent emails, calendar conflicts). Another watches the content pipeline (posting gaps, deadlines). A third tracks sales (invoices, payment signals). They stay silent unless something needs attention. No "all clear" noise, ever.
The system uses multiple AI providers and automatically fails over between them. Cheap models handle triage and data collection. Full-powered models handle reasoning and synthesis. Before any response is sent, a validator strips AI-sounding patterns. The same behavioral corrections apply regardless of which provider generated the response.
Every night, the system evaluates its own performance across system health, output quality, and interaction quality. It finds the worst-performing area, takes a snapshot, attempts a fix, then re-evaluates. If scores improved, the change stays. If not, it rolls back to the snapshot.
End-to-End Message Flow
When a message arrives, it goes through a strict pipeline before the AI ever sees it. Invisible Unicode characters are stripped. A prompt injection scanner checks for manipulation. A security profile is selected based on who is talking. Only then does the AI process the message. Before the response goes out, an output sanitizer strips any exposed secrets, API keys, or file paths. If any step fails, the message is blocked.