The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
Many "cowork / office agents" focus on writing emails, filling spreadsheets, and running workflows. JarvisX-Cowork goes further: building a "Personal AI Creative Assistant" for everyone β when you're out of ideas, it can search the web for references; when you need output, it can take you from copy to visuals all the way through.
Visual creation: text-to-image, image-to-image editing, a refinement workflow from "decent" to "polished" (with professional editing tool integration)
Video capabilities: generate short videos from text/images for product showcases, event teasers, social media content
Real-time information: search web/news/images and provide citable summaries with sources
Dev collaboration: write code, edit code, run commands, organize files (the workflow you know)
The goal isn't just an assistant that "does your office work", but one that "helps you bring ideas to life". From inspiration to output, the fewer tools you switch between, the better.
The table above focuses on "creative capabilities". Beyond this, the project also inherits common baseline capabilities from the Claude Code/Claude Agent ecosystem (files/commands/web/tasks, etc.).
Claude Code (Baseline Capabilities) Overview
| Category | What It Does (Examples) |
|---|---|
| File operations | Read/write/edit files, batch search and locate issues (e.g., Read/Write/Edit/Glob/Grep) |
| Command execution | Run builds, install dependencies, script execution, automation pipelines (e.g., Bash) |
| Web capabilities | Fetch web content for reference retrieval and context (e.g., WebFetch) |
| Task planning | Break down complex goals into steps, continuously push forward and report progress (e.g., Task/Todo) |
| Safety confirmation | Confirm and authorize potentially destructive operations (Human-in-the-loop) |
Python 3.11+ (required when using creative skills)
# Clone the repository
git clone https://github.com/DevAgentForge/JarvisX-Cowork.git
cd JarvisX-Cowork
# Install dependencies
bun install
# Development mode
bun run dev
# Build production version
bun run dist:mac-arm64 # macOS Apple Silicon
bun run dist:mac-x64 # macOS Intel
# Start the app
open "dist/mac-arm64/JarvisX-Cowork.app"
open "dist/mac/JarvisX-Cowork.app"
Troubleshooting Bun Install/Build Issues
If you encounter issues with bun install or bun run dist:* (such as hanging, timeout, or build failures), please refer to the detailed troubleshooting guide:
Option 1: One-click JSON Import: Open App β Settings β Import Config, select configs/jarvisx-config.example.json, then replace the placeholder Keys/URLs with your own.
Option 2: Manual Input: Fill in API Key / Base URL / Model directly on the same settings page, and enable skills and fill in corresponding keys as needed.
To make the experimental environment easier to reproduce, you can use third-party API mirror sites during testing (unofficial, not required).
We make no advertisement or endorsement; please evaluate stability, compliance, and data security risks on your own.
Tip: For production use, prefer official or self-hosted services you trust, keep your keys secure, and avoid committing keys to the repository.
π οΈ Development Guide
# Start development server (hot reload)
bun run dev
# TypeScript compilation
bun run build
# Linting
bun run lint
More details (architecture/modules/skill extensions): docs/summary/PROJECT_ANALYSIS.md
π€ Contributing
The project is continuously being improved. PRs from interested contributors are welcome!
π Acknowledgements
JarvisX-Cowork is a secondary development based on the open-source project Claude-Cowork. Thanks to the original authors and community.
Project origin and inspiration: DevAgentForge/Claude-Cowork.
