Ghast Workbench
Ghast Workbench is the desktop AI agent app for Mac and Windows. Bring multiple models, tools and an Agent Team into one workspace. Coding agents already on your machine, Codex and Claude Code among them, join as teammates.
Chat · Coding · Team
Start with the task and use the workspace that fits it. The guides explain setup, permissions and version-specific controls.
Bring a question and the context it needs. Discuss it with your selected model, then review the reply before using it. Start with an available model configured in Workbench.
Start a conversationSelect a local project, describe the issue and define the allowed changes. Review the diff and actual test output. The task needs access to the project and the runtime required to run its checks.
Follow the tested coding taskGive a named teammate a goal, a delivery and acceptance criteria. Review the returned reply or deliverables, request changes when needed and decide whether to accept the result. Choose the teammates, models and tools the task requires.
Follow the tested Team workflowThe guides record these stable 0.1.33 examples, verified on 19 October 2026 on Apple silicon macOS. Their prerequisites and limits remain part of the walkthroughs.
In a selected project, reproduce a quantity-calculation failure, change only total.py, then inspect the diff and two passing tests. The guide records the model, Python and permission prerequisites; the example does not establish results for other models or platforms.
Using fictional facts, assign a three-row checklist, obtain an independent review, request a revision and manually check the new reply before accepting it. The workflow is text-only and publishes nothing. The guide explains how to handle review findings that still refer to an earlier delivery.
Inspect the Team exampleAgent Team
Bring teammates with complementary specialties around one outcome. Start with the task, pick the teammates and tools it needs, and keep the working context together.
Assign a defined deliverable, follow its status, and review the returned reply or deliverables before accepting the work.
How the team works
A teammate can be asked directly, put in a room with the others, or handed a work item with its own status.
Ask one teammate. Each has its own memory and its own workspace.
Put the same question to everyone, separately. The lead reads the disagreements and turns them into a decision.
@ the owner. The request becomes a work item with a status, picked up in the thread.
Review the returned reply or deliverables against your requirements before accepting the work.
Coding agents already on your machine, Codex and Claude Code among them, join as teammates and keep their own sessions.
Models
Choose from the models available in Workbench. Availability depends on the provider and model configuration.
Bring the model into the same workspace as your conversation, tools, and context.
Enable, test, and select models. Account funds and the model catalog are under 0G Compute settings.
Plugins
Bring tools and skills into Workbench across business, research, development, creative work, and everyday productivity.
Persistent Memory
Persistent Memory keeps useful knowledge and working preferences available as your work continues.
Review, edit, or delete saved memories when needed.
Privacy
A local workspace does not mean every task runs offline. Workbench combines work on your device with configured model providers, tools and optional 0G services. What leaves your device depends on the model, tools, permissions and sync settings you use. The Activity Recorder article explains that context feature. The verifiable execution note is a dated architecture plan, not a live Protocol guarantee.
Prompts, selected context and attachments may be sent to your configured model provider. Supported models are available through 0G Compute.
Workbench encrypts selected backups before uploading them to 0G Storage. Local storage and model requests follow separate data paths.
Coding agents run on your computer, but their model calls and tools can send code or context to external services. Review the provider, permissions and workspace before use.
Choose macOS Apple Silicon, macOS Intel or Windows x64 in the download section. The installed application is labelled Ghast AI; the desktop installers use the Ghast Desktop name.
The desktop application download is free. Model usage and connected services may have separate charges. A free installer does not include free model usage or paid connected services.
Sign in, complete onboarding and configure an available model. Coding also needs access to the selected project and its runtime. Follow the installation guide.
No. Model requests and tools can send selected context to configured services. Review the model, permissions and sync settings in the privacy explanation.
No. It is a separate browser product. The desktop product’s workflows and controls should not be assumed to be available in the extension.
Planned · Ghast Protocol
Through Ghast Protocol, Creators will publish the workflows they run in Workbench as Capabilities that people and agents can buy and invoke. Each would have a price, a scope and measurable delivery commitments, evidence for every order and the Creator’s own capital behind the promise. The method stays private. The Capability Economy describes the market built on these services, and the agentic economy overview sets it in the wider research context.
Ghast Protocol is in development.
Download
Choose the macOS installer for Apple Silicon or Intel, or the Windows x64 installer. The desktop installers use the Ghast Desktop name; the installed application is labelled Ghast AI. The Chrome extension is a separate browser product with a different feature set.
Before your first task, sign in, complete onboarding and configure an available model. Coding also needs access to your selected project and its runtime. Follow the installation guide and check the Workbench release notes. Model usage and connected services may have separate charges.
