Whitepaper / Execution Layer

Ghast Workbench

Ghast Workbench is the live execution layer where users create, run, and coordinate private AI work. It turns objectives into completed work, from a single task to a coordinated team of agents. The knowledge, context, tools, and methods behind each workflow stay under the user’s control.

Workbench runs on macOS and Windows. The desktop installers carry the Ghast Desktop name, and the installed application is labelled Ghast AI. A separate Ghast AI Chrome extension offers a smaller, browser-based feature set; this chapter describes the desktop workspace.

Work happens through three experiences: Chat, Coding, and Agent Team. Models, Plugins, and Persistent Memory support all three.

Chat

Chat is the general-purpose entry point for objective-driven work. An agent interprets the available context, uses permitted tools, and produces a usable output across research, analysis, writing, planning, and operations.

A Chat task starts with the request and whatever the person chooses to include: files, app context and tools. Replies can be regenerated or branched to compare approaches. When a long conversation fills the model’s context window, it can be compacted so the work continues without starting over. Typical Chat work includes summarizing a set of contracts, drafting a market brief from approved sources, or planning a rollout with the relevant documents attached.

Coding

Coding provides a persistent repository- and project-aware environment for software work. It supports understanding, modifying, testing, and reviewing an existing codebase.

Each coding session is tied to a workspace, which is a local project the person selected. The agent reads files, proposes edits and runs commands only within the permissions the person has granted. Every change arrives as a diff that the person reviews before accepting. The project stays on the person’s machine, and only the context a model request needs is sent to the configured provider.

Agent Team

Agent Team coordinates specialized agents around one shared outcome through the operating model described in Ghast Agent Team.

A lead agent owns the result, specialist agents take bounded tasks, handoffs are explicit, and nothing closes until it passes review. A person can pause the team at any checkpoint to add context, redirect the work or approve the next step. Coding agents already running on the user’s machine, such as Codex and Claude Code, can join a team as local agents and keep their own sessions.

Across all three experiences, users can combine specialized knowledge, approved sources, prior context, files, tools, instructions, and guardrails into repeatable private workflows. A repeatable workflow is the raw material for a Capability. Running the same kind of task many times, with the same checks, is how its owner learns what it can reliably promise.

Models

Users select the model for each piece of work and bring it into the same workspace as the conversation, tools, and context.

Models in the desktop app are served through 0G Compute. A person can choose a strong reasoning model for a difficult analysis and a faster, cheaper model for routine background tasks within the same workspace. The model catalog changes as providers add and retire models, and the user guide lists what is currently available and how to fund inference.

Plugins

Tools and skills connect into Workbench across business, research, development, creative work, and everyday productivity.

Plugins, skills and MCP servers extend what a task can do. They also extend what it can reach, so permissions are granted per tool and tool activity is shown as the task runs. A skill can package a repeatable procedure, such as a house style for reports or a checklist for code review, so that every task applies it the same way.

Persistent Memory

Persistent Memory keeps useful knowledge and working preferences available as work continues across tasks. Saved entries can be reviewed, edited, or deleted at any time.

Memory holds facts the person has chosen to keep, such as preferred formats, project conventions and recurring context. Because the person can see and remove every entry, the workspace stays under their control. This matters for the protocol design: in the Privacy Model, private knowledge and memory are part of the method that stays with the Creator.

Where the work goes

A local workspace does not mean every task runs offline. Model requests send the prompt and the selected context to the configured provider, and optional encrypted backups are a separate, opt-in path. The privacy policy and the user guide describe exactly what leaves the device for each feature.

From product usage to paid services

Workbench already powers private AI workflows. The next step is to let people sell them. Creators will offer selected services on Workbench, and through Ghast Protocol, which is in development, publish them as Capabilities that people and agents can buy and invoke.

The path has two stages:

• Private work, live today. People run Chat, Coding and Agent Team in Workbench and pay for the compute their own work consumes.
• Paid services, planned. A Creator who runs the same workflow often enough to state a stable result can offer it as a service. Buyers, whether people or agents, would pay for the delivered result under its Defined Claim, and payment would settle against the evidence.

Payments and Fees sets out how a Buyer would pay and how the proposed fee categories are disclosed. Recognition and Traction records the hackathon result behind the project, and System Overview follows one workflow from Workbench to a Capability.