OpenClacky is an open-source AI agent for coding and general knowledge work. It runs as an interactive terminal session in the current directory or as a local web server with a multi-session chat interface, so you can hold parallel coding, copywriting, and research sessions. Typical tasks range from scaffolding a full-stack project and adding features to explaining an unfamiliar codebase or turning a meeting transcript into assigned tasks.
You supply the model: any OpenAI-compatible endpoint works through a /config step that sets an API key, model name, and base URL, with Claude, GPT, DeepSeek, Kimi, MiniMax, OpenRouter, and OrcaRouter supported without extra setup. Because model choice is yours, you can route code work to one model and delegate subtasks to a cheaper one.
The project's stated design goal is token economy, approached at four layers: sessions that are never restarted, double cache markers, and an insert-then-compress strategy that leaves the system prompt unmutated so compression still reuses the cache; a set of only 16 core tools, with further capability reached through a single invoke_skill meta-tool; background compression of long context while the session is idle, which pre-warms the cache for the next message; and bring-your-own-key model routing.
Relative cost figures published by the project, averaged over internal agent tasks with Claude Code as the 1.0x baseline:
| Agent | Relative cost | Tools |
|---|---|---|
| OpenClacky | ~0.8x | 16 |
| Claude Code | 1.0x | 40+ |
| OpenClaw | ~1.5x | 23 |
| Hermes | ~3x | 52 |
Skills are the extension mechanism. You invoke one with / through browse, fuzzy search, or a direct call, and you can create one by describing it in plain language: the agent drafts a SKILL.md file, breaks the steps down, and runs validation. After each run the agent updates the skill from the execution context and results. Claude Skills, Markdown Pack, and custom formats are supported, and skills can be packaged for sale with encrypted distribution, licence management, and creator-set pricing.
Features
- Two front ends: an interactive CLI in the working directory, and openclacky server for a browser chat UI with parallel sessions
- Bring your own model: any OpenAI-compatible API, with Anthropic, OpenAI, DeepSeek, Moonshot, MiniMax, OpenRouter, and OrcaRouter recognised out of the box
- 16 core tools: a deliberately small tool schema, with everything else reached through the invoke_skill meta-tool
- Cache-oriented session design: persistent sessions, double cache markers, and insert-then-compress so the system prompt is never mutated
- Idle-time compression: long context is compressed and the cache pre-warmed while you are away from the session
- Skill authoring in plain language: describe a skill and the agent drafts its definition, decomposes the steps, and validates the result
- Self-updating skills: each run feeds execution context and results back into the skill definition
- Skill distribution: encrypted packaging, licence management, and creator-set pricing for published skills
- Chat-app integration: connectors for Feishu, WeCom, WeChat, Discord, and Telegram
- Deployment options: desktop installers for macOS and Windows, a one-line shell install, a RubyGems package, and prebuilt container images
