Jarvis Core
A modular assistant core: pluggable skills, an intent planner and a safety policy for risky actions.
What it is
Jarvis Core is a modular core for a personal assistant. Skills plug in as separate modules, a planner turns a sentence into actions, and a policy layer decides what is allowed to run.
What it does
- A FastAPI service with /v1/query, /v1/specs and /v1/health.
- Skills are discovered automatically: any module in the skills package that exposes a SKILL is registered with its tool specs.
- Each tool spec carries a JSON schema for its parameters and a risk level (low, medium, high).
- The planner uses intents and synonyms, and asks a clarifying question when something is missing or ambiguous. Dialogue state is kept per session.
- High-risk actions are blocked by the policy until the user explicitly confirms.
- The planner is rule-based on purpose and can be replaced by tool-calling without touching the skills.
How it is built
Python, FastAPI and Pydantic. The demo below is a TypeScript port of the planner and the policy running in your browser: try asking it to open an app or shut the PC down.
Status
Early core. The Windows connector is a simulation: it reports what it would run and does nothing.
Stack
- Python
Try it
Runs in your browser on sample data.
Sono Jarvis Core. Scrivi 'aiuto' per vedere cosa so fare.
The assistant speaks Italian, like the original. Try:
This is a TypeScript port of the planner and policy from the repository. The Windows connector is a simulation there too: nothing is executed.