OpenClacky is an open-source AI Agent, officially positioned as "the most Token-efficient open-source AI Agent." The project uses the MIT license, with all code public, and is backed by MiraclePlus, ZhenFund, Sequoia China, and Hillhouse Capital, based on the team's two years of accumulated Agentic R&D and harness engineering. It supports BYOK (Bring Your Own Key) and can connect to any OpenAI-compatible model. The official statement claims that under comparable task loads its capabilities are on par with Claude Code, while its cost is significantly lower than other open-source Agents (about 50% of OpenClaw and one-third of Hermes).
[Mechanisms for reducing cost] The official documentation gives four paths: first, an extremely high cache hit rate—sessions do not restart, dual cache markers, Insert-then-Compress, the system prompt is not rewritten, and the cache can still be reused after compression, with a measured cache hit rate approaching 100%; second, a minimal toolset—only 16 core tools are retained, while other capabilities are handled by the Skill ecosystem through a single invoke_skill meta-tool; third, automatic compression during idle time—while the user is away, long context is compressed in the background and the cache is prewarmed, so that the first message after returning directly hits the cache, reducing the cost of the first Token on cold start by more than 50%; fourth, BYOK—the model and cost are chosen by the user, allowing methods such as using Claude to write code and automatically routing subtasks to DeepSeek to further save Tokens.
[Skill system] Skills are described as the core of this Agent. Skills are invoked with /, supporting instant browsing, fuzzy search, and direct invocation; Skills can be created in natural language, with the Agent drafting SKILL.md, breaking down the steps, and executing verification, without writing code; after each run, the Agent updates the Skill based on the execution context and results to achieve self-evolution; it is compatible with Claude Skills, Markdown Pack, and custom formats; polished and mature Skills can be packaged and sold, supporting encrypted distribution, License management, and creator-defined pricing.
[Installation methods] The desktop installer is recommended: on macOS download the .dmg (Apple Silicon / Intel), on Windows download the .exe (Windows 10 2004+ or Windows 11), double-click to installand automatically configure the environment, dependencies, and Skills. It also supports one-line command installation (Mac/Ubuntu) or using Ruby (3.x/4.x), requiring Ruby >= 3.1.0. For more installation instructions, see the official documentation https://www.openclacky.com/docs/installation。镜像在版本标签（v*）发布到 GitHub Container Registry. The container runtime requires --network=host so that the Agent inside the container can access the host Chrome's remote debugging port; macOS / Windows do not support --network=host, so browser automation may be limited. After startup, visit http://localhost:7070 A complete chat interface and multi-session support are available, allowing coding, copywriting, research, and other sessions to run in parallel.
[Configuration and Supported Models] An API Key, Model, and Base URL (any OpenAI-compatible provider) need to be set. Out-of-the-box support includes Claude (Anthropic), GPT (OpenAI), DeepSeek, Kimi (Moonshot), MiniMax, OpenRouter, or any custom endpoint.
[Linux Sandbox Task Limits] A Linux sandbox host can enable task limits via openclacky server --task-cgroup /sys/fs/cgroup/tasks. The host machine must create that cgroup in advance, configure aggregate memory/CPU limits, and enable the memory and cpu subtree controllers. The terminal shell and browser MCP daemon enter the subgroup before exec, while the server remains outside the group; when the subgroup records an OOM kill, terminal results will include resource_exhausted: "memory" and an error, even if the shell masks the failure of the child process. Chrome managed independently also needs to be placed into the same parent group by its launcher. This option is off by default and does not constitute a security boundary against privileged commands.
[Use Cases and Creator Program] As a general-purpose AI programming assistant, it can be used to build full-stack applications, add features, or explore unfamiliar codebases. The official team has also launched a Creator Program: advanced users can turn their workflows into vertical AI experts, enabling encrypted distribution, License management, and self-set pricing, and it has already involved fields such as law, healthcare, and financial planning.
[Limitations and Notes] Version requirement: Ruby >= 3.1.0; on macOS / Windows, --network=host is unavailable, and browser automation may be limited; cgroup task limits are off by default and do not constitute a security boundary. The official team states that the full benchmark report will be published on GitHub; apart from the above information, specific pricing, additional platform support, and performance data are not yet provided.
