What is cockpit-agent?

Cockpit Agent is an edge-cloud collaborative AI agent system designed for smart cockpits, providing a single voice entry point for diverse tasks from millisecond-level vehicle control to multi-day trip planning and deep research. It emphasizes a separation of planning and execution, with LLMs handling understanding and planning while deterministic systems ensure safe and real-time execution of critical functions.

Is cockpit-agent open source?

Yes, licensed under Apache-2.0.

cockpit-agent: Edge-cloud collaborative AI agent system…

Features

01Planning and Execution Separation: LLMs only produce intent/plans, while deterministic executors handle vehicle control with permission validation, ensuring safety.

02Fast/Slow Dual System Architecture: High-frequency, safety-critical commands execute on the edge with millisecond response times and offline availability; complex, cross-domain intents are orchestrated by the cloud LLM Planner.

03Agent Plug-and-Play: New domain agents can be added without modifying core orchestration code, leveraging a unified gRPC contract and Manifest declarations.

04Real Data Priority & Traceability: Integrates with real data sources (e.g., Gaode, Exa, Tushare), rejects mock data in strict mode, and provides source traceability for external data.

05Full-duplex Voice Interaction: Supports a complete streaming voice interaction loop, including KWS, VAD, real-time ASR/TTS, barge-in, and continuous dialogue.

Compatibility

Python

Runtime

Verified via docs

Runtime

Verified via docs

Docker

Deployment

Verified via docs

React

Frontend

Verified via docs

gRPC

Protocol

Verified via docs

Quick start

$ cp .env.example .env

$ make proto

$ make up

Use cases

↳Context-aware Navigation: Planning complex routes with multiple waypoints, real-time ETA, and dynamic re-routing based on ongoing navigation (e.g., "add a gas station, don't be late").

↳Personalized Long-term Memory: The system remembers user preferences (e.g., "wife likes Cantonese food") and applies them to future queries, directly influencing search results rather than just conversational fillers.

↳Mixed Multi-Intent Handling: Simultaneously executing local vehicle controls (e.g., "turn on AC, play music") and sending slower, complex intents to the cloud for parallel processing within a single request.

↳Asynchronous Deep Research: Initiating multi-perspective, iterative background research tasks (e.g., "research solid-state battery progress"), with progress tracking and proactive, referenced reports.

↳Dynamic Scene Creation & Automation: Creating custom car scenarios with natural language (e.g., "create fishing mode: recline seats, dim ambient light"), where LLM compiles during creation, and activation/execution are deterministic.

cockpit-agent

Features

Compatibility

Quick start

Use cases

Alternatives

Comments

Deploy on DigitalOcean