Cortex-M

A cloud-native, modular, and token-efficient personal assistant framework built on open standards — independent of US cloud providers.

---

What is Cortex-M?

Cortex-M is a personal assistant microservice designed to be affordable, stable, highly extensible, and * independent of US cloud providers*. Instead of relying on fragile instruction files and granting the agent broad system access, Cortex-M is built around open standards — using the Model Context Protocol (MCP) for tools and a vector database for memory — keeping token usage lean and behavior predictable. Native support for MistralAI ensures you can run cutting-edge models without vendor lock-in.

The name reflects both its role as the central "brain" of your assistant infrastructure and its foundation on Quarkus — the Supersonic Subatomic Java framework.

---

✨ Features

• Provider-agnostic models — Native support for MistralAI and other non-US providers. Deploy on your own infrastructure with zero vendor lock-in.
• MCP-first tooling — All agent capabilities are defined as structured MCP servers, registered in a database and dynamically loaded at startup or runtime.
• Dynamic tool management — Add or update MCP servers without redeploying the core service.
• Connector architecture — Interact with Cortex-M via any number of lightweight connector services (e.g., Matrix, Discord, Slack), written in any language.
• Multi-connector support — The agent handles multiple simultaneous connectors in parallel, making it ideal for cloud deployments.
• Sandboxed execution — The agent has no access to the local shell or file system; all capabilities are strictly scoped to registered MCP tools.
• Vector memory — A PG-Vector-backed memory store allows the agent to ingest and retrieve experiences as embeddings, replacing token-heavy history files with efficient semantic search.

---

📚 Documentation

See the Wiki for full documentation.

---

🚀 Getting Started

Full setup instructions coming soon.

Prerequisites:

• Docker / Podman
• Java 25+

Quick start:

./mvnw quarkus:dev

Quick start (Docker Compose):

Coming soon!

docker compose up

---

🗺️ Roadmap

• Dynamic MCP server registry (Postgres-backed)
• Multi-connector support via WebSocket + CloudEvents JSON
• Vector memory store (PG-Vector embeddings)
• Direct runtime vector store access by the agent
• Proactive task execution & scheduling (cron-based self-waking)
• Soul / Personality initialization (interactive first-run setup & memory storage)
• Matrix connector
• Telegram connector
• Admin UI with Chat compontent

---

🧠 Design Principles

1. Token efficiency — Context is injected only when needed, not dumped wholesale.
2. Open standards — MCP, CloudEvents, WebSocket; no proprietary lock-in.
3. Cloud-native — Every component runs in a container; state lives in the database.
4. Strict sandboxing — The agent does only what its registered tools allow. Nothing more.
5. Provider independence — Support for non-US models and infrastructure; your data stays under your control.

---

📄 License

GPL3

---

Cortex-M — The central intelligence for your microservice assistant.