OpenAkita

OpenAkita is an open-source, self-evolving AI assistant that coordinates multiple specialized agents to carry out tasks such as web research, file handling, desktop automation and scheduled jobs. It ships as a GUI desktop app (Windows, macOS, Linux) and can also be installed via pip or from source, and it can be reached through chat platforms such as Telegram, Feishu, WeCom, DingTalk and QQ.

What It Is

OpenAkita is a multi-agent AI assistant framework written in Python with a Tauri and React desktop client. You supply an LLM API key, and the assistant plans tasks, calls built-in tools (shell, files, browser, desktop, search, scheduler, MCP) and delivers results in chat. The README states setup is fully GUI-based, with a CLI also available (openakita init, openakita run, openakita serve).

How Multi-Agent and Organization Mode Work

An orchestrator decomposes a request and assigns it to specialized agents that work in parallel, with automatic handoff, failover and a delegation depth limit. Organization orchestration goes further: users design an org chart in the GUI where each node (CEO, CTO, marketing, etc.) is an independent agent with its own identity and memory, sharing information through a three-level blackboard, with message routing, heartbeat checks and deadlock detection.

Memory, Models and Self-Evolution

Memory has two modes: fragment memory with semantic and full-text recall, and an MDRM relational graph covering temporal, causal, entity, action and context links, with automatic routing between them. It supports 30+ LLM providers with failover, including Anthropic, OpenAI, DeepSeek, Gemini and local Ollama or LM Studio (the README notes small local models have limited tool-calling ability). The self-evolution routine performs daily self-checks, root-cause analysis of failures and automatic skill generation or installation.

Extensibility and Security

The plugin system offers eight plugin types, a three-tier permission model and lifecycle hooks, with fault isolation. A skill marketplace and an MCP client (stdio, HTTP, SSE) add further capabilities. A six-layer security model covers path zoning, confirmation gates, command interception, file snapshots, self-protection and OS-level sandboxes (bubblewrap, seatbelt, Windows Mandatory Integrity Control).

Current Status

The latest GitHub release listed is v1.27.42. The source is licensed under AGPL-3.0-only, while brand assets are covered separately by a trademark policy.