How bytebot-ai/bytebot works
Project type and purpose
• Bytebot is a full-stack, containerized microservices application providing an AI-driven virtual desktop agent.
• Key business goal: let an AI “user” with its own Ubuntu/XFCE desktop environment complete tasks via natural-language instructions, exposed through both a Web UI and REST APIs.
File structure
• Top level:
– docker/ & helm/: Docker-Compose files (dev, proxy, core, desktop, Claude) and Kubernetes Helm charts for 5 services (agent, desktop, llm-proxy, ui, postgresql).
– packages/: monorepo of 6 packages:
· bytebot-ui (Next.js frontend & API proxy)
· bytebot-agent & bytebot-agent-cc (NestJS task orchestrators with Prisma migrations)
· bytebot-llm-proxy (LiteLLM service for multi-provider support)
· bytebotd (headless desktop environment + computer-use & input-tracking APIs)
· shared (TypeScript types & utils)
• Each package includes a Dockerfile; NestJS services use Prisma + PostgreSQL; UI uses Next.js server.ts to proxy to NestJS; desktop uses Supervisor + VNC + API modules.
README for additional insights
• The “How It Works” section lists four integrated components: Virtual Desktop, AI Agent, Task Interface, REST APIs.
• Deployment options: Docker Compose for local dev, Helm for Kubernetes.
• Mentions ports: 9991 for task APIs, 9990 for desktop control.
• LLM integration via Anthropic, OpenAI, Google Gemini through the llm-proxy.
Detail
– Include the shared library as a dependency link between UI and agent code.
– Don’t omit the llm-proxy layer or the input-tracking/computer-use modules inside bytebotd.