OpenClacky is a general-purpose AI agent, written in Ruby, built around keeping token spend low. It runs as an interactive terminal agent or as a Web UI with parallel sessions for coding, copywriting, and research, and it connects to any OpenAI-compatible model with your own API key.
Cost reduction comes from four design choices. Sessions never restart and the system prompt is never mutated, so compression still reuses the prompt cache. The agent exposes only 16 core tools and reaches everything else through one meta-tool that invokes Skills. Long context is compressed in the background while the agent is idle, which also pre-warms the cache. Subtasks can route to a cheaper model, for example Claude for code and DeepSeek for subtasks.
The project reports these average relative costs on its internal agent tasks, with Claude Code as the baseline:
| Agent | Relative cost | Core tools |
|---|---|---|
| OpenClacky | about 0.8× | 16 |
| Claude Code | 1.0× | 40+ |
| OpenClaw | about 1.5× | 23 |
| Hermes | about 3× | 52 |
Skills are the agent's main extension mechanism. You browse and call them with a slash command, create new ones by describing them in natural language, and the agent updates a Skill after each run based on its results. Claude Skills and Markdown Pack formats are supported, and a creator program packages Skills for sale with encrypted distribution and license management.
Features
- Bring your own key: any OpenAI-compatible API, with Claude, GPT, DeepSeek, Kimi, MiniMax, OpenRouter, OrcaRouter, and Requesty supported out of the box
- Prompt-cache reuse: an insert-then-compress scheme keeps the system prompt stable, with a reported cache hit rate near 100%
- Minimal tool set: 16 core tools, with further capabilities delegated to Skills through a single meta-tool
- Idle-time compression: compresses long context and pre-warms the cache while you are away
- Subagent routing: routes subtasks to a different, cheaper model
- Self-evolving Skills: create Skills in natural language, and the agent refines each one after it runs
- Skill formats: Claude Skills, Markdown Pack, and custom formats
- Messaging integrations: Feishu, WeCom, WeChat, Discord, and Telegram
- Interfaces: terminal CLI, a Web UI with multi-session support, desktop installers for macOS and Windows, and Docker images
- Coding assistant: scaffolds full-stack apps, adds features, and answers questions about unfamiliar codebases
