Abu Desktop Assistant
Plan and run file, document, and automation tasks on your computer to cut down on manual office work.
- Source repo
- PM-Shawn/Abu-Cowork
- Stars
- ★ 400
- Last updated
- 1mo ago
- License
- NOASSERTION
- Primary language
- TypeScript
- FA score
- 61/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- An individual organizing a desktop or workspace can ask Abu to handle files and run necessary commands.
- Not a fit if
- Users who need official Linux support
- Users unwilling to provide a model API key or configure a local model
- Source review
- 61/100 · Some gaps
What does this agent do, and when should you use it?
Abu is a local-first AI desktop assistant for macOS and Windows, built with Electron, React, and TypeScript. You describe a task in its desktop chat interface; Abu plans steps, invokes tools, reads or writes authorized local files, and runs commands, including to produce documents and reports. It connects to cloud models through Anthropic and OpenAI-compatible protocols, supports local models through Ollama, and can connect to MCP services. Built-in skills, browser and computer control, scheduled tasks, event triggers, and up to five parallel background agents cover both one-off office tasks and recurring workflows. Data and API keys are primarily stored locally, though cloud model and external service calls still require network access.
After you submit a task in chat, Abu plans the work and invokes tools to read or write local files, run commands, and produce PDFs, Office documents, web pages, and reports; Plan Mode can present high-risk steps for confirmation first. It can use the built-in Abu-Browser or optional Chrome extension bridge for web tasks, and Computer Use for guarded screenshots and mouse and keyboard control. Abu supports MCP connectors, web search, and multiple model providers, and can run workflows from cron schedules, file changes, HTTP webhooks, or IM messages. Users can turn multi-step flows into skills and manage conversations, memory, and rules by project.
- An individual organizing a desktop or workspace can ask Abu to handle files and run necessary commands.
- An office worker processing source material can extract tables from PDFs and create spreadsheets, reports, or presentations.
- A team that needs recurring briefings can schedule a cron task to search for news and prepare a digest.
- A user working with an existing signed-in web session can use the built-in browser or Chrome bridge for browser tasks.
- A team that wants chat-driven workflows can connect Lark, DingTalk, WeCom, or Slack and trigger tasks from messages.
How do you install or deploy this agent?
For regular use, download a release package for macOS Apple Silicon, macOS Intel, or Windows x64 from GitHub Releases. The README says official macOS packages are signed and notarized; the Windows installer is current-user only and is not Authenticode-signed.
Building from source requires Node.js 24, npm, Rust stable, and platform build tools. Prepare dependencies and launch the development app:
git clone https://github.com/PM-Shawn/Abu-Cowork.git
cd Abu-Cowork
npm run setup:electron-dev
npm run electron:devThe README identifies Electron as the current development and acceptance target. Tauri source remains for compatibility with already shipped versions, migration, and rollback.
How do you use this agent?
On first launch, open Settings → AI Service Management and select a provider and enter its API key, or install Ollama to discover local models; you can also configure an OpenAI-compatible API URL and key. Verify the connection, choose a model, and start a chat. To use files, a browser, or MCP services, grant file access when prompted and configure the relevant connectors; high-risk operations can require approval.
For example, enter this in the chat composer:
Extract the tables from this PDF and generate an Excel fileYou can also configure scheduled tasks, file watchers, HTTP webhooks, or IM message triggers in Settings.
What are this agent's strengths and limitations?
- Supports Anthropic, OpenAI-compatible endpoints, and Ollama, offering a choice of cloud and local models.
- Provides Request Approval, Smart Review, and Full Autonomy permission modes, and documents plan confirmation for high-risk tasks.
- Combines 29 built-in skills with MCP connectors, browser automation, schedules, and event triggers for desktop workflows.
- The README documents encrypted API key storage and local-first data handling.
- The documented desktop platforms are macOS and Windows; Linux support is not specified.
- Cloud models and external integrations need network access; quick start asks for a provider API key unless using a local model.
- Source builds require Node.js 24, Rust stable, and platform build tools; the docs do not offer a one-command source installation.
- Abu can access local files and run commands, so adopters need to understand and configure its permission modes.
How does this agent compare with similar options?
The README relates Abu to Claude Code's Cowork mode and calls it an open-source alternative to Claude Cowork. It also compares its feature set with regular AI chat and traditional automation.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Abu Desktop Assistant This agent | 61 · Some gaps | Desktop appFree + model costs | ★ 400 | 1mo ago | TypeScript | OpenAI API · Claude API |
| Open Cowork | 65 · Some gaps | Desktop appFree + model costs | ★ 2.2k | 5d ago | TypeScript | Claude Code · OpenAI API · Claude API |
| holaOS | 52 · Major gaps | Desktop appFreemium | ★ 11k | 1mo ago | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| QwenPaw Personal Assistant | 64 · Some gaps | CLIFree + model costs | ★ 36k | today | TypeScript | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The README describes default request approval, plan confirmation for dangerous steps, file and command restrictions, sandboxing, network controls, and encrypted key storage, supporting some least privilege and sensitive-data protection. Full-autonomy mode, broad tools, and external integrations still widen possible effects, while the local-first claim does not fully map data flows to providers, search, MCP, and IM services; these gaps prevent full marks. Security and capability claims come from repository material, but publisher identity and maintainer provenance are not independently established, so attribution evidence is limited.
The repository describes multi-model support, diagnostics, tests, and CI gates, and discloses that Windows tests are currently advisory and non-blocking. This supports ordinary-use reliability, but the supplied material contains no run results, and gives limited detail on concrete error messages or recovery when dependencies are unavailable, so scores remain moderate or low.
The docs identify desktop-office, file, browser, automation, IM, and multi-model scenarios, and describe permission modes, trigger capability levels, supported platforms, and prerequisites. They give less detail on tasks that should not be delegated and on the applicability boundaries of individual integrations.
The README, user and installation guides, build instructions, security policy, and changelog entry point provide a clear information structure. Build prerequisites, version, and Apache-2.0 license are explicit, and the security policy specifies supported versions and reporting channels. The material is feature-heavy, repository-level authorship and maintenance responsibility remain limited, and disclosed limitations such as Windows signing are not comprehensive.
The product claims local file and command execution, planning, skills, projects, previews, scheduling, and multi-model support, indicating usable outputs and added value over ordinary chat. However, these capabilities bring model, integration, and maintenance costs; the material does not quantify benefits or costs, and static evidence cannot establish actual result quality.
README claims receive partial corroboration from package.json scripts, dependencies and version information, plus CI workflows and test structure; the license and security policy also provide concrete inspectable terms. Many feature, security-control, and user-experience claims remain documentation assertions. Static material cannot show that runtime behavior matches those claims, so confidence remains low.
- This desktop agent can read and modify local files, run commands, and connect to external model, MCP, and IM services; understand the selected permission mode and data recipients before use.
- The README claims sandboxing, network isolation, and encrypted key storage, but this static review did not verify their runtime effectiveness.
- The Windows installer is currently not Authenticode-signed; the docs say SmartScreen may require the user to allow it manually.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
✨ Features - Twelve ready-made experts and six ready-made expert teams ship on the Market shelf. Seven new experts — Office documents specialist, Industry researcher, Web designer, QA engineer, Admin assistant, Finance assistant and Contract reviewer — join the existing five, and six teams (Software development, Data analysis, Content creation, Reports & presentations, Finance reconciliation, Recruiting) can be @-mentioned in a conversation right away. The finance assistant handles expense claims, invoices, reconciliation and budget sheets; the contract reviewer walks a contract claus …View on GitHub ↗ -
Release notes excerpt
✨ Features - WeChat now carries images and files both ways. The IM bridge can send images and files out to a WeChat user, shows a "typing…" status while the agent works, and — inbound — a photo and its caption arrive as a single turn instead of two. Inbound images actually reach the vision model now (three root causes fixed), and inbound files and videos are ones the agent can open. Questions and plan-approval requests are relayed to the WeChat user instead of stalling the run. …View on GitHub ↗ -
Release notes excerpt
✨ Features - Images now reach DeepSeek's vision models — The adapter never declared DeepSeek's vision route, so an attached picture was silently swapped for a text placeholder. Vision-capable DeepSeek models now receive the actual image. - One oversized picture can no longer wedge a session — Every provider route now budgets attached-image size and count at admission. Formats no route accepts are re-encoded even when their pixel dimensions are fine — previously a small .bmp walked straight through, entered durable history, and 400'ed every later request in the conversation. The …View on GitHub ↗