Architecture

Manor AI is a monorepo with a Python backend, React frontend, worker runtime, and isolated execution services.

React web app
    |
    v
FastAPI API  <----> PostgreSQL + pgvector
    |                 Redis
    |                 MinIO + JuiceFS
    |                 Ollama (embeddings)
    v
Celery workers (control plane + work queue)
    |
    +---- Sandbox service (per-run containers)
    +---- Integrations, channels, and webhooks

Backend

The API lives in apps/api. It exposes workspace, chat, agent, task, goal, workflow, document, knowledge, integration, channel, scheduling, and operations endpoints.

Shared domain logic lives in packages/core:

• SQLAlchemy models
• services
• agent runtime (the agentic loop, tools, prompt context)
• workflow engine (node-graph runner, importers)
• skills
• permissions
• migrations
• Celery tasks

Frontend

The web app lives in apps/web. It is a React 18 SPA built with Vite and TypeScript. It communicates with the API through the web proxy in Docker and directly with the API during local development.

Worker Runtime

Background execution is split across two Celery workers that share one image:

• Control plane (worker): runs Celery beat, fires scheduled jobs, dispatches work leases, and monitors the deployment. Queue celery.
• Work queue (worker-work): executes long agent plan steps and workflow runs. Queue work.

The split keeps a handful of long-running agent steps from stalling scheduling. Both workers mount the entity filesystem and share the API's codebase and configuration model.

Isolation Boundaries

• The sandbox service executes code in per-run containers with memory/CPU/PID limits — never in the API or worker processes.
• Entity filesystem paths are scoped through the Manor file service on JuiceFS.
• Tool access is constrained by agent settings and HITL governance.
• Data and credentials are Entity-scoped, with Workspace permissions inside each Entity. One self-hosted operator may manage multiple Entities without merging their data or permissions.

Data Stores

Store	Role
PostgreSQL (+pgvector)	Source of truth for structured data and chunk embeddings
Redis	Cache, Celery broker/results, rate limits, JuiceFS metadata
MinIO	Object storage: uploads, artifacts, JuiceFS data blocks
JuiceFS	POSIX entity filesystem assembled from Redis metadata + MinIO blocks
Ollama	Local embedding model (mxbai-embed-large) for document RAG

See Storage for operational detail.