OpenOcta is a Chinese open-source desktop AI agent, distributed as a single Go binary with built-in UI, supporting local memory, MCP/Skills, and multi-channel IM. It runs locally, installs quickly, and provides a desktop-focused experience with offline data options.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
OpenOcta is a desktop-level AI agent project maintained by a Chinese team. It is described as a 100% open-source desktop AI companion with Go-based runtime, a single binary that embeds the Control UI, and features such as MCP, skills, and local memory.
How it works
The project implements Gateway, Agent, and Channels in Go, with a Control UI embedded in the binary. It supports IM channels (WeChat, Enterprise WeChat, DingTalk, Feishu) for remote task commands. It includes a Knowledge Vault local knowledge base and an L4 Evolution mechanism for memory refinement. The front-end UI is delivered as part of the binary and is accessible via the embedded UI.
Getting started
Installation is via official downloads for Windows, macOS (ARM64 and Intel), and Linux (deb/rpm/tar.gz). After installation, the first run initializes the configuration directory. Initial configuration requires connecting to a large language model:
- Way A — Interface configuration (recommended): Open the client Settings -> Model tab, enter API key and select a model.
- Way B — Edit configuration file:
{
"env": {
"vars": {
"MOONSHOT_API_KEY": "sk-你的密钥"
}
},
"agents": {
"defaults": {
"model": { "primary": "moonshot-cn/kimi-k2.5" }
}
}
}
The README shows the path locations for config files on Linux/macOS (~/.openocta/openocta.json) and Windows (%APPDATA%\openocta\openocta.json).
Recent releases
Latest releases include:
- v1.0.6 (2026-07-15): fixes for hardening write_file paths on Windows; preserves sequential assistant/tool pairs when normalizing hi.
- v1.0.5 (2026-07-02): fixes history messages when calling tools; version number update.
- v1.0.0 (2026-07-01): fixes Windows command stage, adds --no-sandbox for root user browser launch, and other minor updates.
- v0.2.8 (2026-06-01): adds support for iFlytek Spark model, NEAR AI Cloud model, Skill online editing, Tool call constraints, and other optimizations.
- v0.2.7 (2026-06-01): enables IM and scheduled tasks with specific digital workers; various optimizations for tool calls, memory persistence, and UI/layout.
Traction
Stars: 3067; Forks: 354; Open issues: 28. (Note: numbers provided as-is.)
Behind the repo
Not provided in the README excerpt beyond general project descriptions; no linked startup/company detail is given.
Caveats
License: Apache-2.0. Project age: created 2026-02-26; last push 2026-08-04. Documentation references Chinese-language docs and a Chinese GitHub landing page. Installation packages are platform-specific (Windows, macOS, Linux). No explicit open issues list is provided here beyond a numeric count.