Skip to main content
This legacy concepts overview remains available by URL. For the canonical concepts docs, use Agent Architecture, Tool System, MCP, and A2A Protocol.
Before diving into the API, it helps to understand how the pieces fit together.

​
The Mental Model

ComponentWhat it doesWhen you need it
AgentLLM-powered reasoning engine that can take actionsAlways - this is your starting point
ToolsFunctions the agent can call to interact with the worldWhen your agent needs to do more than chat
MCP ServerStandardized protocol for exposing tools to any agentWhen you want reusable, shareable tool packages
DeploymentRun your agent as an API endpointWhen you’re ready to integrate with your app

​
Agents

An Agent is an LLM with instructions and capabilities. At minimum, it needs:
  • A name for identification
  • A model (e.g., gpt-4o, claude-sonnet)
  • Instructions defining its behavior
from agentor import Agentor

agent = Agentor(
    name="assistant",
    model="gpt-4o",
    instructions="You help users with their questions."
)
Without tools, an agent can only respond based on its training data. Add tools to let it take action.

​
Tools

Tools are functions your agent can call. They bridge the gap between “knowing” and “doing.” 
from agentor.tools import GetWeatherTool, WebSearchTool

agent = Agentor(
    name="research_assistant",
    model="gpt-5",
    tools=[GetWeatherTool(), WebSearchTool()]
)
Tools can be:
  • Built-in - Celesto provides common tools like weather, web search, and more
  • Custom - Define your own Python functions as tools
  • MCP Servers - Connect to standardized tool servers
Start with built-in tools. Graduate to custom tools or MCP when you need more control.

​
MCP (Model Context Protocol)

MCP is a standard for packaging and sharing tools. Think of it like an API specification for agent capabilities. Why MCP matters:
  • Reusability - Build once, use with any MCP-compatible agent
  • Isolation - Tools run in their own server process
  • Ecosystem - Access tools built by the community
# Connect to an MCP server
agent = Agentor(
    name="gmail_assistant",
    model="gpt-5",
    tools=["celesto/gmail", "celesto/calendar"]
)
You can also build your own MCP servers with LiteMCP.

​
Deployment

Once your agent works locally, deploy it to make it accessible as an API. 
agent.serve(port=8000)
Run on your machine for development.

​
Putting It Together

Here’s the typical progression:
1

Start Simple

Create an agent with just a model and instructions. Test it works.
2

Add Capabilities

Connect tools so your agent can take actions. Start with built-in tools.
3

Build Custom Tools

When built-ins aren’t enough, create custom tools or MCP servers.
4

Deploy

Ship your agent as an API. Start with Celesto Cloud for simplicity.

​
Next Steps

Build an Agent

Deep dive into the Agent API

Add Tools

Learn how to connect tools

Build MCP Server

Create reusable tool packages

Deploy

Ship your agent to production
Last modified on March 4, 2026