Agent Toolkit Desktop

Agent Toolkit Desktop is a standalone Electron workstation over the canonical agent-toolkit serve backend. Its home screen is a semantic world: the workspace is a valley, registered projects are buildings, and shared services have their own recognizable places. The spatial view explains where things belong; the inspectors provide precise control without requiring the user to walk an avatar or run CLI commands.

See the product

These screenshots were captured from the running Electron application with its real backend. They show product behavior; visual approval and remaining composition work are tracked separately in the visual QA log. The larger world and inspector gallery is in the repository README.

Meadow valley	Dusk valley
Project interior	Live Person presence
People roster	Capability Library
Operations	Real PTY terminal

The Library’s Skill catalog also exposes real provenance, prerequisites, compatibility, and a separate Toolkit installation receipt. These close crops keep the detail readable at both window sizes:

Compact	Large

The first-run welcome uses the same pixel sprites as World without implying that example projects or workers already exist:

The current screenshots show real behavior, but do not imply visual sign-off. The visual QA guide records the exact captures opened, remaining composition issues, and what changed after review. The latest world captures include an empty workspace, one project, multiple projects in Meadow and Dusk, and the project interior at compact and large sizes in world-review-2026-10-09/.

Navigate by place or by action

Place	What belongs there	Direct route
World	Workspace context and actual project buildings	/world
Project house	Files and project-scoped actions for that repository	Enter a project from World or Workspace
Library	Reusable skills, Agent Definitions, packs and MCP resources	Library navigation
Archive	Workspace memory and durable records	Memory navigation
Operations	Jobs, loops, swarms, approvals, artifacts and recovery	Operations navigation
Terminal	Real local PTY sessions	Terminal navigation or terminal dock
People	Durable collaborator profiles, separate from live sessions	People navigation
Attention	Failures and items that need a human response	Attention navigation

World objects, menus and the command palette converge on the same typed backend actions. A place should communicate real product state and open a canonical resource; decoration never stands in for a control. See the semantic world contract and UX architecture.

Start using Desktop

Install the platform package from GitHub Releases: AppImage or .deb on Linux, DMG on macOS, or NSIS on Windows. The package includes the V backend and its world assets. No CLI installation, repository checkout, or manual configuration-file editing is required for the basic Desktop setup.

The app guides workspace creation or selection, then exposes capabilities, People, projects, terminals and operations from the interface. A runner must be installed and available before starting a coding session; unavailable runners are reported as such. See packaging for package contents and the current clean-install evidence.

Product model

• AgentDefinition is a reusable Toolkit definition.
• Person is a durable collaborator configured by the user.
• AgentSession is a real live runtime process.
• SwarmRole is a temporary responsibility in one swarm run.

An offline Person remains visible in People and is absent from the world. Only a live session, job or run with real project context can create runtime presence. The Person remains after the process exits. The People guide documents supported configuration and current session limits.

Current status and evidence

The executable coverage source is workflows.yaml. It marks each user journey ok, partial, blocked, or not-implemented with evidence; read it before assuming a workflow is complete. Visual quality is tracked separately in VISUAL_QA.md. Person sessions now have durable V lifecycle records and recover truthfully as interrupted when the local PTY is gone. The Start review now preloads the saved goal into a one-time task, previews the role/goal/task prompt, and sends it through documented interactive options for Claude, Codex, OpenCode and Copilot. Cursor and Muse open without an initial prompt and offer a copy action. The prompt itself is not retained in the PersonSession record.

The table below is a user-facing summary of that ledger. “Partial” means the GUI path exists but a documented end-to-end capability or recovery case is still missing; it does not mean the user should substitute a CLI command.

Workflow	Desktop entry point	Status	Current limit
First-run workspace creation	Onboarding	Ready	Creates or reuses a workspace through the GUI; a real supervised-backend crash/restart is verified before entering the world.
Switch and inspect workspaces	Settings → Workspace	Ready	The GUI rejects a missing destination without changing the active workspace, switches to a second workspace, links a project there, verifies its World building, then switches back without changing the original workspace. Backend recovery after a successful switch is exercised as part of the same flow.
Enter a project and open its real files	World project house, Workspace	Ready	Files are read from the linked project through the typed backend.
Discover and install capabilities	Library	Partial	A shared search filters Skills and Agent Definitions. Skills show typed provenance, declared requirements and target compatibility separately from receipt-backed install evidence. Workspace packs are not browsable, compiled Copilot Skills and Agent Definitions are outside installation, and Agent Definition compatibility is not declared here.
Configure and probe MCP	Library → MCP	Ready	The probe checks the local executable, not a live model connection.
Start and inspect jobs	Operations; World presence	Ready	Only backend-confirmed active work appears in the World.
Use and recover a real terminal	Terminal or terminal dock	Ready	PTY lifecycle is real and covered through close, exit and restart recovery.
Create, edit, archive and import People	People	Ready	Munder import is reviewed and inert; saving never starts a process.
Start a Person	People → Start	Partial	One-time prompts work for Claude, Codex, OpenCode and Copilot; no durable conversation/resume exists. Token/cost limits and non-inherited isolation are not enforced by this interactive PTY.
Review swarm recipe and topology	Operations → Swarms	Ready	Start configuration previews real roles, policies, gates and budgets.
Bind People to swarm roles	Operations → Swarms → People	Partial	Preferences resolve before start, but role processes are not canonical PersonSessions and do not appear as People in the World.
Run and observe a swarm	Operations → Swarms	Partial	Run inspection and controls exist; adapter-backed workers, handoffs and failure recovery need an end-to-end Desktop run.
Run/report and schedule loops	Operations → Loops	Ready	GUI supports one-shot run/report plus schedule preview, create, disable and recovery.
Find and launch core actions	Command palette (Ctrl/Cmd+K)	Ready	Palette actions enter the same GUI review flows and canonical backend operations.

For exact evidence and the scope of each status, use the journey record in workflows.yaml; this summary deliberately does not upgrade any ledger result.

This guide is maintained against the shipping Electron + React application; the retired native-V GUI is not part of the product. V remains the canonical engine, server and CLI. Older captures and migration records live in VISUAL_QA_HISTORY.md and are evidence, not current art direction. The README gallery uses real application screenshots rather than illustrative SVG mockups. CI, license and release destinations are plain text links. Dated comparison proposals are indexed in design-notes and are not current feature requirements.

For architecture and contribution commands, see the Desktop package guide, design contract, and contributor guide.