Skip to content
/

Open Claude Cowork

open-claude-cowork · DevAgentForge/Open-Claude-Cowork · ★ 3.3k · last commit 2026-03-21

Primitive shape
No installable primitives
00

Summary

Open Claude Cowork — Summary

Open Claude Cowork (also called "Agent Cowork") is a minimal Electron desktop application built on the Anthropic Claude Agent SDK. It presents a single-window GUI for creating and executing autonomous tasks via Claude Code, designed as an open-source alternative to "Claude Cowork." Users create tasks and choose execution paths; the desktop app handles OAuth, renders tool call outputs visually, and manages task context — without requiring any terminal, Claude Code installation, or development environment knowledge.

The project is notable for using @anthropic-ai/claude-agent-sdk directly (v0.2.6) and SQLite via better-sqlite3 for task state. With 3,290 stars and minimal codebase, it is the simplest desktop wrapper in this batch. Compared to Aperant or Eigent, it has no parallel agents, no kanban, and no methodology — just a GUI surface for individual task execution. Closest to claude-conductor (seed) in philosophy (minimal scaffolding, GUI-first) but implemented as a desktop app rather than a markdown template.

01

Overview

Open Claude Cowork — Overview

Origin

Created by DevAgentForge organization. 3,290 stars, no external contributors visible. Version 0.0.2-0.1.0. TypeScript + Electron. No license specified in repo root.

Philosophy

From README: "Agent Cowork is an open-source alternative to Claude Cowork — a desktop AI assistant that helps with programming, file management, and any task you can describe."

Key positioning:

  • "Not just a GUI. A real AI collaboration partner."
  • "No need to learn the Claude Agent SDK — just create tasks and choose execution paths."
  • Reuses existing ~/.claude/settings.json without modification

The core thesis: Claude Code is powerful but terminal-only. Agent Cowork gives it a native desktop app surface with visual feedback for complex tasks and easy task tracking.

MiniMax Partner Note

The README includes a sponsored section for MiniMax M2.7 — a frontier LLM positioned as an alternative model for agentic execution. The project appears open to multi-model use.

Technical Approach

The app patches @anthropic-ai/claude-agent-sdk@0.2.6 (noted in package.json patchedDependencies) — suggesting it customizes the SDK behavior rather than using it off-the-shelf.

02

Architecture

Open Claude Cowork — Architecture

Distribution

  • Type: Electron desktop application
  • Platforms: macOS (arm64, x64), Linux, Windows
  • Install: Download from GitHub Releases or bun run dev from source

Directory Tree

Open-Claude-Cowork/
├── src/
│   ├── electron/         # Electron main process
│   └── ui/               # React frontend
├── electron-builder.json # Packaging config
├── vite.config.ts
└── package.json

Required Runtime (Build from source)

  • Bun or Node.js 22+
  • Claude Code (installed and authenticated for OAuth)

Config Files

  • ~/.claude/settings.json (reused from Claude Code — not modified)
  • SQLite database (better-sqlite3, internal to app)

Target AI Tools

Claude Code (Anthropic Claude Agent SDK). Potentially MiniMax (sponsor mention, not confirmed in code).

Key Dependencies

  • @anthropic-ai/claude-agent-sdk@^0.2.6 (patched)
  • better-sqlite3@^12.6.0 — task state storage
  • Electron 39+
  • React 19 + Vite + Tailwind CSS
  • Zustand (state management)
  • react-markdown + rehype-highlight (rendering tool outputs)
03

Components

Open Claude Cowork — Components

Desktop UI

View Purpose
Task creation panel Describe task; choose execution path
Tool output renderer Visual display of Claude Agent SDK tool calls/results
Task history SQLite-backed task log

Core Components

Component Purpose
Electron main process App lifecycle, window management
Claude Agent SDK integration Task execution via @anthropic-ai/claude-agent-sdk
SQLite task store Persist task history via better-sqlite3
React UI Rendering, markdown output, Tailwind styling

CLI

None. No user-facing CLI binary.

Skills / Commands / Hooks

None. The app reuses ~/.claude/settings.json but adds no Claude Code commands, skills, or hooks of its own.

05

Prompts

Open Claude Cowork — Prompts

Open Claude Cowork has minimal prompt content. The application passes the user's task description directly to the Claude Agent SDK; no CLAUDE.md, SKILL.md, or methodology files are present in the repository.

Verbatim excerpt 1 — README task framing

Agent Cowork is an open-source alternative to Claude Cowork — a desktop AI assistant
that helps with programming, file management, and any task you can describe.

Not just a GUI.
A real AI collaboration partner.
No need to learn the Claude Agent SDK — just create tasks and choose execution paths.

Prompting technique: Direct task-to-API pass-through. The user's natural language task description becomes the agent's prompt with no transformation. No spec format, no phases, no methodology injection.

Verbatim excerpt 2 — Settings reuse

Reuses your existing ~/.claude/settings.json
No development environment or Claude Code installation required.

Prompting technique: Inherited context. The app reuses Claude Code's existing settings (CLAUDE.md contents, permissions, etc.) without adding its own instruction layer — a zero-prompt-injection approach.

Note

Because the app patches @anthropic-ai/claude-agent-sdk@0.2.6, there may be custom prompt injection inside the patched SDK, but this is not visible from the public repository.

09

Uniqueness

Open Claude Cowork — Uniqueness & Positioning

differs_from_seeds

Open Claude Cowork is philosophically closest to claude-conductor (seed) — both prioritize minimal scaffolding over methodology, letting the underlying agent (Claude Code or Claude Agent SDK) do the work. But where claude-conductor ships 15 markdown templates as the value proposition, Open Claude Cowork ships an Electron desktop GUI. Among the 11 seeds, no seed ships a desktop application; Open Claude Cowork is the minimal viable expression of "what if Claude Code had a window." Compared to other desktop apps in this batch (Eigent with CAMEL-AI backend, Aperant with 12-agent kanban, Multica with web app + squads), Open Claude Cowork is the simplest: single-agent, single-task, no methodology, no channels, no worktrees. Its key differentiator within the batch is using @anthropic-ai/claude-agent-sdk directly (the only framework in the batch to do so) rather than spawning Claude Code as a subprocess.

Positioning

  • Lowest complexity in the batch — intentionally minimal.
  • Direct Claude Agent SDK integration (v0.2.6) — not a subprocess wrapper.
  • "Open-source alternative to Claude Cowork" explicitly positioned against Anthropic's own product.

Observable Failure Modes

  1. SDK version pinning: Patched @anthropic-ai/claude-agent-sdk@0.2.6 — SDK updates break the patch.
  2. No error recovery: No retry logic, no memory, no crash recovery documented.
  3. Single agent: Cannot parallelize work; sequential only.
  4. No license: No LICENSE file — legal reuse unclear.
  5. Minimal maintenance signal: No contributors listed; last commit 2026-03-21.
04

Workflow

Open Claude Cowork — Workflow

Phase Flow

Phase Description Artifact
1. Task Describe User types task in GUI Task description
2. Execution Path Select User chooses how to execute Execution config
3. Agent Execute Claude Agent SDK executes task Tool call stream
4. Output Render Visual rendering of tool calls and results Task output
5. History Task saved to SQLite Task record

Approval Gates

None explicit. The Claude Agent SDK may prompt for confirmation on dangerous operations depending on OAuth settings.

Simplicity Note

Open Claude Cowork is the simplest framework in this batch. It has no phases beyond "describe task + execute." No parallel agents, no kanban, no channels, no methodology.

06

Memory Context

Open Claude Cowork — Memory & Context

Memory Type

SQLite (better-sqlite3) for task history. No cross-session agent memory.

Memory Persistence

Project-level (app-local SQLite).

State Files

  • Internal SQLite database (app data directory, path not documented)

Cross-Session Handoff

Task history accessible via SQLite but no explicit handoff mechanism.

Context Compaction

Inherits Claude Agent SDK defaults.

07

Orchestration

Open Claude Cowork — Orchestration

Multi-Agent

No. Single agent per task. No parallelism.

Execution Mode

Interactive-loop. User submits task; agent executes; results displayed.

Multi-Model

No. Claude Agent SDK (Anthropic) only.

Isolation

None.

Orchestration Pattern

None (single agent, sequential).

08

Ui Cli Surface

Open Claude Cowork — UI / CLI Surface

CLI Binary

  • Exists: No
  • Binary name: null

Local UI Surface

Desktop Application

  • Type: Electron desktop application
  • Platforms: macOS (arm64 + x64), Linux, Windows
  • Tech stack: Electron 39 + React 19 + Vite + Tailwind CSS
  • SDK: @anthropic-ai/claude-agent-sdk@0.2.6 (patched)

Features

  • Task creation panel (natural language task input)
  • Tool call output renderer (markdown + code highlighting via rehype-highlight)
  • Task history (SQLite-backed)
  • Execution path selection
  • Reuses ~/.claude/settings.json (inherits Claude Code permissions/config)

Comparison with CLAUDE-FLOW UI

Claude-flow has no local UI. Open Claude Cowork is arguably the simplest possible desktop wrapper for Claude — no parallel agents, no kanban, no methodology, just a GUI window that passes task descriptions to the Anthropic SDK and renders the results. This is the minimum viable desktop cowork surface, compared to Aperant (12 agents, kanban, QA loop) or Multica (full web app with squads, autopilots, skills) at the more complex end.

Observability

  • Task history via SQLite (internal)
  • No audit log, no streaming event feed, no cost tracking

Related frameworks

same archetype · same primary tool · same memory type

Goose (Block/AAIF) ★ 46k

General-purpose AI agent (not just code) with security-first tool inspection, recipe-based shareable configurations, and 15+ LLM…

Vibe Kanban ★ 27k

Eliminate the overhead of planning, switching between agent terminals, and reviewing diffs by providing a single web dashboard…

1Code ★ 5.5k

Cursor-like desktop experience for Claude Code and Codex with cloud background agents, event-driven automations, and a full…

Crystal (stravu) ★ 3.1k

Manage multiple parallel AI coding sessions in isolated git worktrees from a single desktop GUI.

Maestro (RunMaestro) ★ 3.0k

Orchestrate unlimited parallel AI agent sessions with a keyboard-first desktop app including Group Chat coordination and Auto Run…

AgentsMesh ★ 2.1k

Multi-tenant workforce platform that gives every team member a squad of AI coding agents coordinated through channels, pod…