# Agentor

A Python framework and API for building and serving long-running AI agents, with durable, resumable runs, agent-to-agent communication, and pluggable tools.

Agentor is an open-source Python framework for building and deploying AI agents that keep running over time. Agents can persist, resume, and fork their runs, communicate with each other, and expose a tool-search interface that reduces how much tool context is loaded at once. A decorator-based MCP server implementation is built on FastAPI.

Optional extras add tools with heavier dependencies, such as Gmail and Calendar integration, GitHub, Slack, Postgres, and web scraping, while the base install stays small. Agentor works with any OpenAI-compatible provider, including OpenRouter, Groq, Together, Fireworks, DeepSeek, vLLM, Ollama, and Anthropic's and Gemini's compatible endpoints, so a model or provider can be swapped with one line of code. An agent can also be served as an API with a single method call.

Website: https://docs.celesto.ai/agentor