Agentao

Agentao is an embeddable agent runtime for Python. A developer imports it into an application and gets a stateful, tool using assistant running in process, and the same runtime can also run as a command line tool or as an Agent Client Protocol server, so hosts written in other languages can drive it without reimplementing anything. The project describes it as local first and governed, with permissions, MCP support, memory and audit replay built in. It is for backend engineers adding agent features to a product, developers building IDE or editor plugins, and teams that need controlled, auditable agent behaviour. Python hosts use an in process SDK with callbacks, while IDEs and Node, Go or Rust hosts talk to it over stdio with NDJSON and JSON RPC 2.0. Built in tools cover files, shell, web and search, alongside an MCP client, a permission engine, multi tenant working directories, conversation compression, SQLite backed memory and switching the model at runtime. Six extension points allow custom tools, skills, MCP servers, permission rules, memory and sandbox policy, and the CLI offers slash commands, plan mode, memory and replay. Documentation is published in English and Simplified Chinese, with blueprints for common product shapes and chapters for security reviewers and operations staff. Agentao is free and open source under the MIT licence, with no paid plan or pricing page. Model costs depend on the LLM provider you configure.