An agentic AI orchestration platform that manages LLM-powered agents, tool execution, task queues, projects, knowledge bases and visual workflows — all through a modern web UI backed by a streaming FastAPI server.

Make your own workflow tailored to your work

flowchart LR
    React[React UI] <--> Orchestrator

    Orchestrator <--> Local[Local Store]

    Local <--> Knowledge[Knowledge]
    Local <--> Conversations[Conversations]

    Orchestrator <--> Worker["Worker(s)"]

    Worker --> LLM[LLM Backend]
    Worker --> Tools[Tool Executor] <--> Sandbox[Sandbox]

System

acai/
├── orchestrator/       # FastAPI server, chat, agents, projects, config, dispatcher
├── worker/             # Worker HTTP app, LLM abstraction, sandbox backends
├── tasks/              # Task graphs: converse, think, uber, scribe, dynamic workflows
├── tools/              # Agent tools: filesystem, shell, git, code, web, search, …
├── queue/              # SQLAlchemy-backed work queue (WorkQueue)
├── agents/             # Built-in agent definitions (system.j2 templates)
├── models/             # Model backends: vLLM, llama.cpp, SGLang, TensorRT-LLM, …
├── scheduler/          # Multi-provider scheduling
├── cli/                # CLI entry points: orchestrator, worker, mcp
├── ui/                 # React frontend (Vite, Chakra UI v3, React Flow)
│   └── src/
│       ├── components/ # Pages & reusable components
│       ├── services/   # api.ts (REST/SSE), types.ts
│       ├── contexts/   # WebSocketContext (socket.io)
│       └── layout/     # Shell layout with sidebar
└── plugins/            # Optional plugin discovery

Local Data (Json, git)

workspace/              # Runtime data (default location, configurable)
├── conversations/      # Global conversation histories
├── projects/           # Per-project data, tasks, conversations
├── workflows/          # User-defined workflow graphs
├── knowledge/          # Markdown knowledge documents
├── agents/             # User agent overrides (copy-on-write)
└── work.db             # SQLite task queue database

The orchestrator is a FastAPI application augmented with python-socketio for real-time events. It owns:

• ChatStore — filesystem-backed conversation persistence with a three-tier layout: global (workspace/conversations/), project-scoped (workspace/projects/<project>/conversations/), and task-scoped (workspace/projects/<project>/<task_id>/).
• AgentStore — agent definitions with Jinja2 system-prompt templates, tool lists, sandbox flags. Built-in agents ship in acai/agents/; user overrides are written to workspace/agents/ with copy-on-write semantics.
• WorkQueue — a SQLAlchemy task queue (SQLite by default) with statuses pending → curating → ready → in_progress → completed/failed/review, dependency tracking, priority ordering, and timeout reaping.
• ProjectStore — project metadata (definition.json), scaffolding, and git-clone support.
• KnowledgeStore — slug-based Markdown documents organized as subject / subsubject / title.
• ProviderScheduler — routes LLM calls across configured providers.
• LoadBalancer — selects healthy workers for dispatch.

Workers are lightweight HTTP servers that register with the orchestrator and expose two core endpoints:

• POST /worker/llm/complete — streams tokens back as SSE (token, reasoning, tool_call_delta, done, error).
• Tool routes — execute agent tools in-process or forward sandboxed calls through the SandboxProxy.

LLM backends are pluggable via a create_llm factory: vLLM, or a generic OpenAI-compatible endpoint.

When an agent is marked uses_sandbox, tool calls flagged sandbox=True run inside an isolated environment. Supported backends:

Inside the sandbox, acai mcp starts a minimal HTTP tool server that the worker's SandboxProxy communicates with.

Every chat interaction runs through a TaskGraph — a directed execution graph that orchestrates the agent loop:

Visual workflows are built in the React Flow-powered editor and persisted as JSON node/edge specs.

The UI is a React 19 + TypeScript single-page application built with Vite and styled with Chakra UI v3. Communication with the orchestrator uses three channels:

Key pages: Home, Conversations, Projects (Kanban board), Workflows (visual graph editor), Knowledge, Agents, Tasks, Settings, Status.