Abu: A Local-First AI Desktop Agent with Self-Evolving Skills

What Abu Is and What It Does on Your Machine

Abu is a desktop application, not a web service. It runs as an Electron app on the user's machine, reads and writes local files, executes commands, and completes tasks autonomously. The README positions it as inspired by Claude Code's Cowork mode: the user describes what they need, and Abu figures out the steps, invokes tools, and reports back.

The README's comparison table identifies the key differences from both regular AI chat and traditional automation. Unlike a chat interface, Abu can plan and execute multi-step tasks end-to-end. Unlike traditional automation, it accepts natural language instructions rather than requiring scripted workflows. Unlike either, it stores all data locally: the README states 100% local data, privacy-safe as a key property.

Version 0.50.0 includes a workspace file tree with a code canvas that lets users browse, preview, and edit files in a side panel alongside the conversation. The declarative progress panel lets the AI model declare its own plan steps and status via a report_plan mechanism, so users can see what Abu is working on without reading raw output. Inline visualization widgets render charts, HTML, and Mermaid diagrams directly inside the chat.

Skills: 29 Built-In Plus Self-Evolution

Abu ships with 29 built-in skills. Skills in Abu are reusable patterns that encode specific multi-step workflows. A skill can be invoked by the agent when a user's request matches a known pattern, which speeds up common tasks and improves consistency across sessions.

The self-evolving feature is the distinctive mechanism. After a user runs a complex multi-step flow, Abu proactively offers to crystallize that flow into a new skill. This means the skill library grows from actual usage patterns rather than requiring manual definition. The result is a skill set that reflects the specific workflows of the individual user or team over time.

Custom skills can also be written manually and added to the library. The builtin-skills/ directory in the repository holds the default skill definitions. The README also mentions MCP connectors: one-click integration with Playwright, GitHub, and other tools that extend what Abu can invoke.

The skills/ directory in the top-level entries is separate from builtin-skills/, indicating a separation between the bundled defaults and the user's own skill library. Projects in Abu can have their own skill and MCP configuration, isolated from other projects.

Plan Mode, Permission Control, and the Safety Sandbox

Abu distinguishes between read-only operations and operations that write, delete, send, or install. High-risk operations trigger Plan Mode: Abu presents a step-by-step plan and waits for the user to click Confirm before proceeding. Only read-only operations run while Abu is waiting for approval.

Permission control operates at three levels: Request Approval (all tool use requires confirmation), Smart Review (Abu reviews its own actions for risk before proceeding), and Full Autonomy (no confirmation required). The permission level can be set per conversation without affecting other conversations.

The README also describes a content safety scan that checks agents, skills, and memory for prompt injection and dangerous instructions. This is a hardening layer on top of the permission model, designed to detect cases where an external source attempts to manipulate Abu's behavior through its inputs.

The security sandbox uses a seatbelt mechanism with network isolation. The README names these in the features table without detailing their implementation. For users who are concerned about what commands Abu executes on their machine, the Request Approval permission mode is the authoritative control: no tool call runs without explicit user confirmation in that mode.

Multi-Agent Parallel Execution and IM Integration

Abu can run up to 5 background agents simultaneously, with real-time progress displayed in the interface. This allows independent subtasks to run in parallel: for example, writing documentation in one agent while another runs tests in a second. The progress panel updates as each agent completes its declared steps.

IM integration is a notable feature for teams in organizations that use Lark, DingTalk, WeCom, or Slack. Abu can be configured as an IM bot that receives messages, processes them as agent tasks, and responds in the channel. The README shows this as a supported capability in the features table, distinguishing Abu from desktop AI assistants that operate only through their own interface.

Scheduled tasks use a cron-based scheduling mechanism. HTTP endpoints, file changes, and IM messages can trigger agent tasks automatically. For recurring workflows like daily report generation or automated monitoring, this removes the need for manual invocation.

The AI service management panel supports multiple providers with health checks. The multi-endpoint provider presets in version 0.50.0 include Volcengine, Bailian, and Zhipu as curated presets alongside a unified add/edit modal for custom endpoints. Per-model capabilities (vision, tools, reasoning, token limits) are declared per model, so Abu knows what each configured model can do.

Limitations: Desktop-Only and No Shared Memory

Abu runs as a desktop application on macOS and Windows (Electron-based). It is not a web application, not a browser extension, and not a server-side process. Teams that need a shared AI agent accessible from multiple machines or by multiple users simultaneously will find that Abu's local-first design does not fit that use case.

The personal memory system remembers preferences and work habits, but the README describes this as personal memory, implying it is scoped to the individual user's installation. There is no documented mechanism for sharing Abu's memory or skill library across a team.

The README describes Abu as an alternative to Claude Code's Cowork mode, but the two tools serve different purposes. Claude Code is a coding assistant that operates inside a terminal and reads code directly in the repository context. Abu is a general-purpose desktop agent for producing reports, generating documents, running commands, and interacting with IM channels. A developer who wants AI assistance inside their editor or terminal should use a coding tool like Claude Code. A developer who wants an autonomous agent for broader desktop and communication tasks is Abu's intended user.

The DISCLAIMER.md in the top-level entries signals that the project treats liability and usage boundaries as first-class documentation concerns, which is worth reading before deploying Abu in a professional context.

License, Maintenance, and Build Information

The package.json in the repository lists the license as Apache-2.0, and the README displays an Apache-2.0 badge. However, the GitHub repository metadata marks the license field as NOASSERTION. Before redistributing Abu or incorporating it into a product, reading the LICENSE file in the repository directly and consulting the FORKING.md documentation is advisable.

The last push was on 2026-09-28, and the most recent release is v0.50.0 from 2026-09-16. The project is in active development. The ELECTRON-NEAR-TERM-ROADMAP.md and ELECTRON-TRANSITION-RELEASE.md files in the top-level entries indicate that the Electron build path has been actively evolved and has a documented near-term roadmap.

The project ships signed and notarized macOS builds, which the README highlights as a recent improvement in v0.50.0. Windows builds are also available. The RELEASING.md and RELEASE-CHECKLIST.md files indicate that the release process is documented and maintained.

Editorial conclusion

Abu is a practical fit for developers and teams who want an autonomous AI agent that stays on the local machine, supports Chinese enterprise IM tools like Lark, DingTalk, and WeCom, and can grow its own skills from real workflow patterns. It is not a cloud service, not a coding assistant IDE extension, and not a replacement for Claude Code in a developer workflow where the AI works directly inside a repository. Before installing, verify the license: the package.json shows Apache-2.0, but the GitHub repository metadata marks the license as NOASSERTION. Check the LICENSE file in the repository directly before redistributing or building a product on top of Abu.

Frequently asked questions

Official sources

1. Issues
2. PM-Shawn/Abu-Cowork on GitHub
3. Project website
4. README
5. Releases