Cheshire Cat
A self-hosted Python AI agent framework with a web chat UI, Ollama and vLLM support, cloud model connections, and a GPL-3.0 license.
Cheshire Cat is a self-hosted Python framework for people learning how AI agents work or building custom assistants for research and creative projects. It pairs a local web chat interface with an API, so you can test an agent in conversation and use it inside another application. It's open source under GPL-3.0.
The framework runs on your own machine and connects to Ollama or vLLM for local model use. It also supports OpenAI, Anthropic, Gemini and OpenRouter. Model choice matters for privacy: the framework runs locally, while cloud model connections send requests to the selected provider.
Its design centers on agents, callable tools, directives and hooks. Agents handle conversations and tool calls; directives add behaviors such as document retrieval, memory and guardrails. Hooks let developers inspect or modify incoming tasks and outgoing replies. Python plugins can add these behaviors alongside custom API endpoints, models and authentication handlers.
Multiple agents can share a chat, and you can choose an agent for each message or let agents call one another. Native Model Context Protocol (MCP) support connects agents to servers that supply tools, prompts and resources. The built-in interface supports multiple chats, while AG-UI streaming exposes generated text, tool calls and agent events as they happen. For applications that need their own interface, the REST API supports agent messaging and custom routes with role-based authentication.
Version 2 is an unstable alpha with breaking changes and incomplete features. The maintainers do not recommend it for production use yet.
