agentor

Fastest way to build and deploy long-running AI agents—with durability, observability, and security.

Docs | Examples

Features

🚅 Quick Start

Installation

The recommended method of installing agentor is with pip from PyPI.

pip install agentor

This installs the v0.1.0 line, built on Agentor's own agent engine — durable, forkable runs included.

Tools with heavy or vendor-specific dependencies ship as extras, so the base install stays small:

pip install "agentor[google]"   # GmailTool, CalendarTool
pip install "agentor[all]"      # every optional tool

Available extras: google, exa, git, github, slack, postgres, scrapegraph, all.

pip install git+https://github.com/celestoai/agentor@main

Build and Serve an Agent

Build an Agent, connect external tools or MCP Server and serve as an API in just a few lines of code:

from agentor.tools import GetWeatherTool
from agentor import Agentor

agent = Agentor(
    name="Weather Agent",
    model="gpt-5-mini",  # Use any LLM provider - gemini/gemini-2.5-pro or anthropic/claude-3.5
    tools=[GetWeatherTool()]
)
result = agent.run("What is the weather in London?")  # Run the Agent
print(result)

# Serve Agent with a single line of code
agent.serve()

Any OpenAI-compatible provider

Point base_url at any provider that speaks OpenAI's /chat/completions — OpenRouter, Groq, Together, Fireworks, DeepSeek, vLLM, Ollama, or Anthropic's and Gemini's compatible endpoints — with no extra dependency:

agent = Agentor(
    name="Assistant",
    model="openrouter/auto",
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)

Run the following command to query the Agent server:

curl -X 'POST' \
  'http://localhost:8000/chat' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "input": "What is the weather in London?"
}'

agent.serve() gives you an ordinary ASGI app, so host it wherever you already run Python services.

Durable Runs

Give the agent a store and every run becomes an append-only event log that survives process death. Resume an interrupted run, or fork any persisted run — even a completed one — into a new, independent run that keeps the full trace, the model's reasoning included when the provider returns it:

from agentor import Agentor
from agentor.engine.store import FileStore

agent = Agentor(name="Assistant", model="gpt-5-mini", store=FileStore("runs"))

result = agent.run("Draft a launch plan")   # persisted under result.run_id
# Interrupted mid-run? agent.resume(result.run_id) picks up where it left off.

fork = agent.fork(result.run_id, "Make it punchier")  # fork.run_id is a new run; parent untouched

fork and resume have async twins, afork and aresume.

Tracing

Tracing is off unless you ask for it. A trace carries prompts, tool arguments, tool results, and the model's reasoning when a provider returns it — so nothing leaves your process by default.

Turn it on for an agent:

agent = Agentor(name="Assistant", enable_tracing=True)   # needs CELESTO_API_KEY

Or decide per run:

agent.run("public question")
agent.run("contains customer data", tracing=False)  # this run sends nothing
agent.run("debug this", tracing=True)               # trace just this one

tracing= is accepted by run, arun, chat and stream_chat. View traces at celesto.ai/observe.

Agent Skills

Skills are folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks.

Agent Skills help agents pull just the right context from simple Markdown files. The agent first sees only a skill’s name and short description. When the task matches, it loads the rest of SKILL.md, follows the steps, and can call a shell environment to run the commands the skill points to.

• Starts light: discover skills by name/description only
• Loads on demand: pull full instructions from SKILL.md when relevant
• Executes safely: run skill-driven commands in an isolated shell

Skill layout example:

example-skill/
├── SKILL.md        # required instructions + metadata
├── scripts/        # optional helpers the agent can call
├── assets/         # optional templates/resources
└── references/     # optional docs or checklists

Using a skill to create a GIF:

from agentor.tools import ShellTool
from agentor import Agentor

agent = Agentor(
    name="Assistant",
    model="gemini/gemini-3-flash-preview",
    instructions="Your job is to create GIFs. Lean on the shell tool and any available skills.",
    skills=[".skills/slack-gif-creator"],
    tools=[ShellTool()],
)

async for chunk in await agent.chat("produce a cat gif", stream=True):
    print(chunk)

Create an Agent from Markdown

Bootstrap an Agent directly from a markdown file with metadata for name, tools, model, and temperature:

---
name: WeatherBot
tools: [get_weather]
model: gpt-4o-mini
temperature: 0.3
---
You are a concise weather assistant.

Load it with:

from agentor import Agentor

agent = Agentor.from_md("agent.md")
result = agent.run("Weather in Paris?")

Build a custom MCP Server with LiteMCP

Agentor enables you to build a custom MCP Server using LiteMCP. You can run it inside a FastAPI application or as a standalone MCP server.

Agentor uses the official MCP Python SDK 2.3 or newer within the 2.x line. FastMCP is not an Agentor dependency; the LiteMCP decorator API is unchanged.

from agentor.mcp import LiteMCP, get_token

mcp = LiteMCP(name="my-server", version="1.0.0")

@mcp.tool(description="Get weather for a given location")
def get_weather(location: str) -> str:

    # *********** Control authentication ***********
    token = get_token()
    if token != "SOME_SECRET":
        return "Not authorized"

    return f"Weather in {location}: Sunny, 72°F"

mcp.serve()

LiteMCP vs FastMCP

Key Difference: LiteMCP is a native ASGI app that integrates directly with FastAPI using standard patterns. FastMCP requires mounting as a sub-application, diverging from standard FastAPI primitives.

📖 Learn more

Connect an Agent to an MCP Server

Use MCPServer for a remote server that supports Streamable HTTP. Agentor discovers its tools, including paginated results, and closes the connection after the run:

from agentor import Agentor
from agentor.mcp import MCPServer

agent = Agentor(
    name="Assistant",
    model="gpt-5-mini",
    tools=[MCPServer(url="https://your-server.example/mcp")],
)

Hosts that supply their own HTTP transport can pass http_client_factory to MCPServer. The factory receives headers and timeout keyword arguments and must return a new httpx2.AsyncClient. Agentor owns that client's lifecycle, closing it when connection setup fails or the MCP connection closes.

Search the web through MCP

The Parallel Search example connects an agent to web search and page extraction without a Parallel API key. It also includes a tools-only mode that runs without model credentials:

uv run python examples/parallel_search.py --tools-only "Agentor durable agent runs"

Agent-to-Agent (A2A) Protocol

The A2A Protocol defines standard specifications for agent communication and message formatting, enabling seamless interoperability between different AI agents.

Key Features:

• Standard Communication: JSON-RPC based messaging with support for both streaming and non-streaming responses
• Agent Discovery: Automatic agent card generation at /.well-known/agent-card.json describing agent capabilities, skills, and endpoints
• Rich Interactions: Built-in support for tasks, status updates, and artifact sharing between agents

Agentor makes it easy to serve any agent as an A2A protocol.

from agentor import Agentor

agent = Agentor(
    name="Weather Agent",
    model="gpt-5-mini",
    tools=["get_weather"],
)

# Serve agent with A2A protocol enabled automatically
agent.serve(port=8000)
# Agent card available at: http://localhost:8000/.well-known/agent-card.json

Any agent served with agent.serve() automatically becomes A2A-compatible with standardized endpoints for message sending, streaming, and task management.

📖 Learn more

🤝 Contributing

We'd love your help making Agentor even better! Please read our Contributing Guidelines and Code of Conduct.

📄 License

Apache 2.0 License - see LICENSE for details.