Release Candidate

Magic Key
Assistant

Local-first operations harness for small teams, combining AI-assisted chat, durable records, retrieval, planning, and administrative continuity.

Project Summary

Magic Key Assistant is a local-first operational continuity platform for micro-teams that need AI support without surrendering their working memory to a cloud SaaS product. The system turns everyday conversation, project notes, research, and follow-up into durable operational records, including actions, decisions, blockers, source links, knowledge gaps, and recurring work streams. Its upcoming release combines multi-backend LLM routing, self-correcting retrieval, formal state machines, trust controls, and a FastAPI admin dashboard into a practical assistant for teams that need continuity, not another chat window.

Core Capabilities

Multi-Backend AI Routing Routes work across Ollama, OpenAI-compatible endpoints, OpenRouter, LM Studio, vLLM, llama.cpp, and Anthropic, with configurable model roles for speed, critique, and synthesis.
Local Complexity Classifier Uses a small local model to classify requests before generation, sending simple tasks to faster settings and complex work to higher-quality pipelines.
Self-Correcting Retrieval Combines HyDE retrieval, entity extraction, sub-question decomposition, synonym expansion, chunk enrichment, and knowledge-health checks to improve answers from imperfect internal material.
Operational Records Stores actions, decisions, blockers, and source links as typed records with validated state transitions rather than loose notes or ad-hoc task lists.
Continuity & Pulse Dashboard Surfaces overdue actions, stale blockers, unresolved decisions, planning collisions, pending reviews, and knowledge gaps in an urgency-sorted operational feed.
Trust-Gated Automation Applies quiet hours, change gates, noise budgets, audit logging, and confirmation requirements so autonomous suggestions remain controlled and reviewable.

Technical Implementation

Python 3.12+ FastAPI SQLite WAL ChromaDB Discord.py Tailwind CSS Ollama llama.cpp Docker Windows Installer

The release version uses 27 sequential database migrations, a service container for lazy-instantiated shared services, request tracing for per-stage timing and model usage, circuit breakers for service resilience, encrypted backup support, and hardware-aware setup that can recommend local model configurations based on CPU, RAM, disk, and GPU capacity.

System Design

  • Application-layer AI: Designed to consume multiple LLM backends rather than depend on a single hosted provider or proprietary workflow.
  • Prompt-compatible tool use: Uses a backend-agnostic XML-style tool-calling protocol, allowing agentic workflows across local and hosted models.
  • Symbolic safeguards: Uses formal state transitions, causal tracing, planning rules, review queues, and confirmation gates around operational work.
  • Admin-first deployment: Provides a web dashboard for system health, model routing, review queues, retrieval logs, continuity views, settings, activity, and knowledge management.

Impact & Outcomes

  • Reduced operational drift: Recurring work, unresolved decisions, stale blockers, and overdue actions are captured and surfaced instead of disappearing into chat history.
  • Durable institutional memory: Conversations, source links, artifacts, records, retrieval feedback, and knowledge gaps become a searchable operating layer for the team.
  • Local-first AI adoption: Teams can use local or hosted models according to cost, privacy, speed, and quality needs without rebuilding the workflow around each backend.
  • Practical autonomy: The system nudges, routes, reviews, traces, and escalates work while preserving human control over consequential changes.
Screenshot of a message from Magic Key Assistant app indicating memory has been stored and added to the memos folder, with a preview of strategic and operational focus for MTR Magic Key.
Screenshot of a Discord chat conversation discussing the technical features and functionalities of an AI assistant called Magic Key Assistant, with multiple users and messages.