Abu-Cowork: Run an AI Desktop Assistant with Electron

Overview

Abu is an Electron desktop office assistant inspired by Claude Code’s Cowork mode. It uses a dual-protocol adapter for Anthropic and OpenAI-compatible APIs to plan and perform work on your computer, such as reading and editing files, running commands, and creating reports. It is designed for macOS and Windows, with local models available through Ollama as well as cloud providers.

Rather than only answering questions, Abu can carry out multi-step tasks through tools, scheduled workflows, browser automation, and integrations. Its local-first approach keeps data and API keys on the user's machine, while permission controls ask for approval for consequential operations. This makes it most relevant to people who want an AI assistant to act on desktop work, not just draft responses in a chat window.

Key Features

• Agent planning and tool use for file operations, commands, and document creation, with a plan-and-confirm mode for high-risk steps.
• Connects to Anthropic and OpenAI-compatible services, supports Ollama discovery, and allows custom endpoints.
• Provides built-in skills for office documents, design, browser automation, and other workflows, with support for custom and self-evolving skills.
• Runs up to five background agents in parallel and can group conversations, skills, and connectors into projects.
• Supports scheduled tasks and event triggers, including cron, file changes, HTTP webhooks, and IM messages.
• Offers MCP connectors, browser automation, and computer control, with permission settings and safety checks.
• Integrates with Lark, DingTalk, WeCom, Slack, and D-Chat for team chat workflows.

Use Cases

• An office worker can ask Abu to extract PDF tables into a spreadsheet or draft a presentation, rather than manually moving content between apps.
• A developer or analyst can delegate local file review, command execution, and report generation to an agent, using approval controls for sensitive steps.
• A small team can use an IM integration to invoke tasks from a supported chat platform and set workspace or user restrictions.
• Someone with recurring research or reporting work can schedule a workflow and have Abu produce a digest or other deliverable.

Getting Started

Download a macOS or Windows installer from the GitHub Releases page, then configure an AI service in Settings. For a local model, install Ollama; Abu can discover local models. See the README and Installation Guide for details.

Alternatives

• UI-TARS-desktop: UI-TARS focuses on operating apps through visual computer control, while Abu plans and carries out broader file, command, and document tasks.
• craft-agents-oss: Craft Agents centers on organizing customizable agent sessions and connected services, while Abu focuses on completing desktop file, command, and document tasks.
• Kun: Kun provides a shared desktop and terminal workspace for varied agent work, while Abu is focused on planning and carrying out tasks on the computer.
• goose: Goose is a general-purpose agent offered as a desktop app, CLI, and API, while Abu is an Electron assistant focused on desktop task execution.

Considerations

The repository is active, with frequent commits and releases, but measured contribution is highly concentrated in one person. That concentration is a continuity risk for a desktop agent with many integrations and security-sensitive capabilities.

Abu can read and modify files, run commands, and control a computer. Its safeguards are not equivalent across platforms: the README describes macOS Seatbelt confinement for shell commands, while Windows lacks an OS-enforced file-access sandbox. Network isolation also relies on proxy configuration rather than OS enforcement. Evaluate these boundaries against your data and threat model before enabling broad permissions.

Building from source requires Node.js 24, Rust stable, and platform build tools. The project is licensed under Apache-2.0; its README notes that the Enterprise edition requires a separate license.