Mini Agent is a single-agent demo project that shows the execution pipeline behind agents built on MiniMax's M2.5 model. It runs a complete agent loop over a basic toolset for file-system and shell operations, and reaches the model through an Anthropic-compatible API that supports interleaved thinking for long, multi-step tasks.
The runtime keeps working past a single context window. A session note tool stores key information so the agent can pick a task back up across sessions, and conversation history is summarised automatically once it approaches a configurable token limit. Every request, response, and tool execution is logged for debugging.
Tools come from two sources: a bundled set of 15 skills covering documents, design, testing, and development, and Model Context Protocol (MCP) servers for capabilities such as knowledge-graph access and web search. Configuration is a single YAML file holding the API key, API base, model name, step limit, and workspace directory; the global platform uses https://api.minimax.io and the China platform https://api.minimaxi.com.
The project also speaks the Agent Communication Protocol (ACP), so the agent can be registered as an agent server in an editor such as Zed and driven from the editor's agent panel. Tests cover the tool classes and LLM client, the session note tool and MCP loading, and end-to-end agent execution.
Features
- Full agent execution loop: a complete run cycle with built-in file-system and shell tools
- Persistent memory: a session note tool carries key information across separate sessions
- Context management: conversation history is summarised automatically up to a configurable token limit
- Skills library: 15 bundled skills for documents, design, testing, and development
- MCP tools: native Model Context Protocol support for knowledge-graph access and web search
- ACP integration: registers as an agent server in editors such as Zed
- YAML configuration: API key, API base, model, maximum steps, and workspace directory in one file
- Logging: detailed records of every request, response, and tool execution
