caiao 1.0.1

CAIAO — Server-as-atomic-unit framework for tool orchestration

CAIAO

Server-as-atomic-unit framework for tool orchestration.

CAIAO turns every solver, tool, or external capability into an independent Server process. The Hub discovers, routes, and manages them. Add a new capability by writing one Server file and registering it — no core code changes.

---

Two paths, same contract

	MCP SDK Path	Lightweight Path
What	pip install "git+https://github.com/LaobaiAi/CAIAO.git#subdirectory=caiao"	Copy 3 files into your project
Dependencies	mcp>=1.0.0	Python stdlib only
Async	Full asyncio lifecycle	Synchronous, in-process or subprocess
Best for	Production platforms, team projects, AI ecosystem	Embedded Python (Abaqus/Blender), prototypes, CLI tools
Upgrade path	—	Replace ~10 lines to switch to MCP SDK

---

Quickstart (MCP SDK path)

pip install "git+https://github.com/LaobaiAi/CAIAO.git#subdirectory=caiao"

from caiao import CAIAOClientHub
from caiao.discovery import discover_server_configs

configs = discover_server_configs(servers_dir="./servers")
hub = CAIAOClientHub(configs)
await hub.start_all()
result = await hub.call_tool("my_tool", {"param": "value"})

Create a new project:

caiao init my-project
cd my-project
caiao new server my-solver
# → servers/my-solver/server.py + caiao.yaml

---

Quickstart (Lightweight path)

Copy caiao_lightweight/server.py, hub.py, subprocess.py into your project.

from server import CAIAOServer, tool
from hub import Hub

class MyServer(CAIAOServer):
    name = "my_server"

    @tool
    def greet(self, args: dict) -> dict:
        return {"hello": args.get("name", "world")}

hub = Hub()
hub.register(MyServer())
result = hub.call_tool("greet", {"name": "CAIAO"})

---

Project structure (after caiao init)

my-project/
├── servers/                # All CAIAO Servers live here
│   └── my-solver/
│       ├── server.py       # MCP Server with list_tools/call_tool
│       └── caiao.yaml      # Declarative manifest
├── gateway/                # Application entry point (optional)
│   └── main.py             # Hub initialization + API
├── caiao/                  # (lightweight only) Framework files
├── caiao.yaml              # Project-level config
└── requirements.txt

---

CLI commands

caiao init <name>            Create a new CAIAO project
  --lightweight              Use zero-dependency lightweight path
caiao new server <name>      Scaffold a new Server
  --kind atomic-mcp          Default: standard Server
  --kind composite           Pipeline orchestration Server
caiao validate               Validate all caiao.yaml manifests
caiao list                   List all Servers with status
caiao doctor                 Check environment (Python, MCP SDK, etc.)

---

How it works

User → Application → Hub.call_tool("analyze", {...})
                         │
                         ├─→ Local handler? ──→ Execute in-process
                         ├─→ Tool registry? ──→ Route to Server subprocess
                         ├─→ Lazy? ───────────→ Start Server on-demand
                         └─→ Not found? ──────→ Semantic fuzzy match

Each Server is an independent process communicating via stdio JSON-RPC (MCP protocol). The Hub manages the full lifecycle: start, stop, restart, health check, metrics.

---

Exit mechanism

CAIAO is designed to be removable. Your Servers are standard MCP servers — they import from mcp.server, not from caiao.

To remove CAIAO:

1. pip uninstall caiao (or delete caiao_lightweight/ directory)
2. Replace Hub calls in your application with direct MCP Session connections
3. Your Servers continue to work unchanged — each is a standalone MCP server

No server code changes required. The framework is a convenience layer, not a hard dependency.

---

Documentation

Full documentation at caiao-knowledge-base/ (this repository):

• Core concepts and design principles
• Server patterns and reusable code patterns
• Development guides for both implementation paths
• Protocol specification v1.0
• Operations manuals (merge, test, review, absorb)

---

Requirements

• Python >= 3.11
• MCP SDK path: pip install mcp
• Lightweight path: Python stdlib only