Isolated Linux desktop workspaces for AI agents.
Server defines 25 tools with consistent naming (verb_noun pattern: workspace_*, profile_*, viewer_*, mcp_control_*) and descriptions present for all. However, parameter schemas are incomplete: most tools lack detailed type information, constraints, and descriptions for individual parameters. Input schemas show only parameter names and basic descriptions (e.g., workspace_exec has 'command' and 'args' but no type constraints or format guidance). Output schemas are not documented in the visible code. Error handling is present (enforce_agent_mutation checks, recovery_hints) but lacks actionable guidance for LLMs. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications in metadata.
Get the current MCP control mode and permission state.
Update the MCP control mode (active/paused/locked) with user confirmation.
Validate and check a workspace profile for issues.
Delete a workspace profile.
Get details of a specific workspace profile.
List all available workspace profiles.
Open a workspace from a profile with optional setup and startup applications.
Parameter schemas lack type constraints and format guidance. workspace_exec 'command' and 'args' have descriptions but no type definitions (string vs array element types). workspace_input 'input_type' lacks enum constraint (keyboard|mouse). No min/max bounds on numeric or array parameters.
Output schemas not documented in visible code. Tools return IpcResponse, ProfilePutResult, WorkspaceBrowserTargets, etc., but LLMs cannot see field definitions, types, or which fields are always present vs optional. Breaks downstream tool chaining.
Tool annotations missing despite clear risk classifications. workspace_stop (WRITE), profile_delete (DESTRUCTIVE), workspace_exec (WRITE) lack readOnlyHint/destructiveHint/idempotentHint annotations. LLMs cannot infer safety properties from names alone.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Create or update a workspace profile.
Close one or more workspace viewers.
List all open workspace viewers.
Open the workspace viewer GUI.
Click on an element in the browser.
Navigate to a URL in the browser.
Open a URL in a browser within the workspace.
Search for elements in the browser DOM.
Capture a snapshot of the current browser state including DOM and visual content.
List available browser targets in a workspace.
Run diagnostic checks on the workspace environment and return a structured report.
Execute a command in a running workspace and return the output.
Return machine-readable guardrails for isolated workspace actions, including concurrency limits, permission ceilings, and approval requirements.
Send keyboard or mouse input to a workspace.
Capture a screenshot of the workspace desktop environment.
Start a new isolated workspace with specified configuration, environment variables, and optional profile setup.
Get the current status of a workspace, including running processes, resource usage, and agent mode.
Stop a running workspace and clean up its resources.
Error handling lacks actionable recovery guidance. enforce_agent_mutation() returns generic 'MCP live control is paused' without suggesting next steps. No error categorization (retryable vs user-fixable vs fatal). Recovery hints exist but are not consistently applied to all error paths.
Parameter descriptions are generic and lack format/constraint details. 'Workspace ID (defaults to DEFAULT_WORKSPACE_ID if not provided)' does not specify format (alphanumeric, max length, allowed characters). 'Profile configuration object' for profile_put lacks schema reference or field documentation.