Loading ToolBench
macOS desktop app to monitor and manage running Claude Code sessions
This is a Tauri desktop application, not an MCP server. It has no MCP protocol implementation, no transport mechanism for remote clients, and tools are exposed as Tauri IPC commands rather than MCP tools. The codebase shows 6 command handlers for managing Claude Code sessions, but they lack proper schema definitions, comprehensive descriptions, and error handling patterns required for agent integration. Tools are tightly coupled to Tauri's native APIs and desktop UI concerns, making them unsuitable for agent use. The application is domain-specific (session management UI) rather than a general-purpose tool server.
Focus the terminal containing a specific session
Get all active Claude Code sessions
Kill a Claude Code process by PID
Register a global keyboard shortcut to toggle the window
Unregister the current global keyboard shortcut
Update the tray icon title with session counts
Not an MCP server, this is a Tauri desktop application with no MCP protocol implementation. Tools are Tauri IPC commands, not MCP-compatible tool definitions.
Input schemas are missing or inferred only from parameter lists. No explicit JSON Schema definitions visible for any tool. The rubric requires explicit schema presence; inferred schemas cap tool score at 50.
Tool descriptions are extremely brief (under 20 characters in most cases). 'Get all active Claude Code sessions' (34 chars) lacks guidance on when to call it or what the output structure is. Descriptions must be 50-200 chars and answer: what, when, and what does it return.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 27 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 33 | - | v1 |
No error handling guidance. Tools like 'kill_session' (destructive) and 'focus_session' (may fail if PID invalid or process already terminated) provide no recovery instructions. Agents need to know what to do on failure.
Destructive operations (kill_session) lack confirmation or dry-run capability. Agents should be able to preview impact before executing irreversible actions.
Parameters lack descriptions in most cases. 'pid' and 'project_path' in focus_session have descriptions, but most other parameters are under-documented. LLMs cannot infer meaning from type alone.
Output schemas not documented. Agents don't know what fields 'get_all_sessions' returns (session IDs, names, statuses?). Without documented outputs, downstream tool composition breaks.
Tool 'update_tray_title' accepts Tauri-specific types ('tauri::AppHandle') as parameters. This is not portable across contexts and breaks the agent abstraction, Tauri app handles are not agent-accessible resources.
No permission gates or scope declarations. 'kill_session' is destructive but has no permission check. Agents should declare required permissions and verify authorization before executing.
Parameter formats and constraints not documented. 'pid' is u32 but no guidance on valid ranges. 'shortcut' string format is undefined (what keyboard shortcut syntax is accepted?). LLMs need explicit constraints.