This repository provides a detailed analysis of the Kimi agent architecture (K2.5 model), offering insights into its design and operational principles for AI researchers and engineers. It focuses on Kimi's innovative shift towards skill-based environments, enabling models to function as operating system users rather than mere API consumers, thereby enhancing agentic capabilities.

How It Works

Kimi's architecture represents a departure from traditional tool-use models. Instead of relying on discrete APIs, it furnishes the AI with general-purpose computing contexts, including persistent filesystems, browser automation capabilities, and process execution environments. This design allows the Kimi model to operate with the autonomy of an OS user, directly interacting with and manipulating its environment.

Quick Start & Requirements

This repository serves as an analytical resource rather than a runnable application. To understand specific aspects:

• Architecture Overview: Refer to analysis/how-kimi-works.md.
• Terminology: Consult GLOSSARY.md.
• Agent Types: Compare prompts/base-chat.md with prompts/ok-computer.md.
• Skill Definitions: Examine source-code/skills/docx/SKILL.md.
• Technical Details: Explore deep-dives/README.md.

Highlighted Details

• Analysis of Moonshot AI's Kimi agent architecture (K2.5 model).
• Key architectural shift from tool-use to skill-based environment interaction.
• Model functions as an OS user, leveraging filesystems, browser automation, and process execution.
• Extracted artifacts include system prompts for six agent types, skill definitions for DOCX, XLSX, PDF, and WebApp outputs, schemas for 38 distinct tools, and runtime source code.
• Methodology relies on extraction via standard public interfaces, without adversarial techniques or decompilation.

Maintenance & Community

No specific details regarding maintainers, community channels, or ongoing development are provided in the README.

Licensing & Compatibility

The repository features mixed licensing: extracted materials (prompts, source code, tool schemas) are under CC0 1.0, while original analysis and documentation are under CC BY 4.0. Third-party components retain their upstream licenses. The analysis itself is compatible with commercial use, but the licensing of extracted Kimi artifacts should be considered.

Limitations & Caveats

The analysis was conducted using Claude Code and is limited to information exposed through Kimi's standard public interfaces. No binaries were decompiled, and no authentication mechanisms were bypassed. The findings represent the state of the system as observed through conversational queries.