An open-source, enhanced version of Claude Cowork built specifically for programmers.
Explore the docs »
Getting Started
·
View Demo
·
Report Bug
·
Request Feature
·
Contact Us
Table of Contents
CrewBench is a desktop UI and control layer for AI coding agents such as Claude Code, Gemini CLI, and Codex.
It provides visibility, control, and rollback over how AI agents read and modify a real codebase.
Note
CrewBench is built on top of AionUi (Apache-2.0) and inherits all upstream AionUi capabilities, while adding developer-focused observability and guardrails.
| Feature | Description |
|---|---|
| Universal Model Support | Work with cloud and local models through a unified interface, including OpenAI-compatible APIs and local runtimes. |
| Agent Coworking Environment | Multiple agents and tasks can run in parallel with isolated context, enabling real cowork-style workflows. |
| Direct File System Access | Agents operate directly on real project files via a file-tree interface, not abstract chat outputs. |
| File Timeline | Every file change is recorded in a chronological timeline, making agent activity inspectable and reviewable. |
| Action Logs & Audit Trail | Full logs of both AI and human actions provide observability, debugging capability, and reproducibility. |
| Live Preview & Editing | Generated artifacts are surfaced as editable, previewable files rather than hidden chat responses. |
| Extensible & Customizable UI | The interface is developer-controlled, supporting layout flexibility and custom CSS theming. |
CrewBench adds AI-aware file version tracking, recording every agent-initiated change:
- Full content snapshots per change
- Line-based diffs between versions
- Operation type (
create,write,delete) - Agent attribution
- Timestamped history per workspace
This complements Git by tracking agent behavior, not commits.
Action logs provide a comprehensive audit trail of all actions performed by agents and humans in the workspace. Each log entry includes:
- Timestamp
- Action type (create, write, delete)
- Agent attribution
- File path
- Content snapshots
- Line-based diffs
| Focus Area | What It Covers | Why It Matters |
|---|---|---|
| Observability | • File timelines and diffs • Clear agent vs human attribution • Action-level history |
Lets developers see exactly what happened, who did it (human or AI), and in what order |
| Control / Guardrails | • Path-based read/write restrictions • Operation-level permissions • Confirmation gates for risky actions |
Prevents agents from touching the wrong files or making unsafe changes |
| Reproducibility | • Inspectable agent action sequences • Debugging and diagnosis of AI behavior |
Makes AI behavior debuggable, auditable, and repeatable |
| Human + Agent Workflows | • Continuous history across human and agent edits • No context loss between steps |
Enables seamless collaboration without losing state or intent |
AionUi provides the base platform, including:
- Multi-agent orchestration
- Desktop UI infrastructure
- File system access
- Local-first execution
- Cross-platform support
CrewBench builds on top of this foundation and focuses specifically on software engineering workflows.
Note
- Does not replace Git
- Does not replace an IDE
- Does not introduce new AI models
- Does not focus on prompt presets
CrewBench operates at the control and observability layer, not the editor or model layer.
- macOS: 10.15 or higher
- Windows: Windows 10 or higher
- Linux: Ubuntu 18.04+ / Debian 10+ / Fedora 32+
- Memory: Recommended 4GB or more
- Storage: At least 500MB available space
- CrewBench is a desktop application and does not require any prerequisites.
- Download the latest release from here.
- Install the downloaded file.
- Now you can run the
CrewBench as an Applicationfile.
Cowork With CLI AI Agents (Unified Developer Interface)
CrewBench acts as a graphical cowork layer on top of existing command-line AI coding tools like Gemini CLI, Claude Code, Codex, Qwen Code, Goose, and more.
| Capability | Description |
|---|---|
| Multi-Agent Sessions | Run multiple AI coding agents simultaneously |
| Unified Interface | One UI instead of multiple terminals |
| Context Isolation | Each agent and task runs with independent memory |
| Local Execution | Agents run locally on your machine |
Cowork Directly on Your Codebase
Agents operate on real project files — not abstract chat outputs.
| Developer Need | How CrewBench Supports It |
|---|---|
| Code Exploration | File-tree based navigation |
| File Modification | Agents read and write actual project files |
| Batch Operations | Rename, merge, and reorganize files |
| Safe Changes | All changes are tracked and reviewable |
Inspect Agent Output, Not Just Results
CrewBench surfaces agent outputs as first-class artifacts you can inspect and modify.
| Visibility Feature | Description |
|---|---|
| Multi-Format Preview | Code, Markdown, HTML, documents, images, diffs |
| Live File Sync | Previews update automatically when files change |
| Inline Editing | Edit or refine agent output directly |
| Change Awareness | See what changed, not just the final state |
File Timeline and Action Logs
Engineering-grade observability for AI coding agents.
| Observability Tool | Purpose |
|---|---|
| File Timeline | Chronological history of every file change |
| Action Logs | Full audit trail of agent and human actions |
| Change Attribution | Clear mapping of who (or which agent) did what |
| Reproducibility | Inspect and debug agent behavior |
Parallel Task Execution Without Context Collision
Run multiple tasks simultaneously without cross-contamination.
| Workflow Pattern | Support |
|---|---|
| Parallel Sessions | Multiple active tasks at the same time |
| Isolated Memory | No cross-task confusion |
| Long-Running Work | Persistent sessions for ongoing work |
Remote Coworking and Team Workflows
Control and observe agents remotely while keeping execution local.
| Capability | Description |
|---|---|
| WebUI Access | Control CrewBench from other devices on the network |
| Local Agent Control | Agents continue running on the host machine |
| Server Deployment | Suitable for self-hosted setups |
| CrewBench Teams | Collaborative multi-developer workflows (coming soon) |
Developer-Controlled Interface
CrewBench adapts to your workflow — not the other way around.
| Customization | Description |
|---|---|
| CSS Customization | Full control over UI appearance |
| Layout Flexibility | Shape the interface around your workflow |
| Personal Setup | Make CrewBench feel like your own IDE |
- [Completed] File Timeline
- [In Progress] Action Logs
- [Not Started] GuardRails
To contribute to CrewBench, please read the CONTRIBUTING.md file.
Distributed under the project_license. See LICENSE.txt for more information.
Abhinav Sharma - x/abhinavsharma_x
Project Link: https://github.com/CrewBench/CrewBench
- Electron — for enabling cross-platform desktop application development
- Node.js — for the underlying runtime and tooling ecosystem
- SQLite — for reliable, local-first data storage
- OpenAI — for OpenAI-compatible APIs and ecosystem standards
- Google Gemini — for multimodal and large-context model support
- Anthropic Claude — for advancing safe and capable coding agents