Loading ToolBench
MCP server exposing Apple Silicon sandbox engine to Claude Code sessions. Provides one-shot and persistent sandbox execution with three isolation tiers (Seatbelt/macOS, MicroVM/Linux, native process), plus desktop browser automation.
21 tools with mostly complete schemas and descriptions, but significant gaps in parameter documentation and error handling guidance. Tool names follow verb_noun convention well (sandbox_run, session_create, desktop_click). Descriptions average ~100-150 chars, meeting baseline minimums but lacking LLM-optimized context about WHEN to use each tool vs alternatives. Parameters have types and defaults, but many lack constraint documentation (e.g., timeout ranges, allowed tier values). No tool descriptions mention error recovery paths or prerequisites. Output schemas are not documented, LLMs cannot predict response structure for chaining. Desktop automation tools (desktop_click, desktop_type, desktop_key) lack coordinate validation guidance.
Extract text content from the current page using CDP.
Navigate Chromium to a URL in the desktop VM.
Click at coordinates on the desktop VM's virtual display.
Press a key or key combination on the desktop VM.
Capture a screenshot of the desktop VM's virtual display.
Scroll the desktop VM's virtual display.
No output schemas documented. LLMs cannot predict response structure (fields, types, pagination) for downstream tool chaining. E.g., session_list returns what? Array of session objects? With which fields?
Destructive tools (session_destroy, desktop_session_destroy) lack confirmation/dry-run guidance. No error recovery hints. Descriptions do not warn about irreversibility.
Parameter constraints underdocumented. timeout (1-600s), tier enum (A|B|C), ttl_seconds (60-86400), memory_gb (1-8), cpus (1-8) lack explicit range/enum declarations in descriptions. LLMs may pass invalid values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Create a persistent desktop VM session with Xvfb + Chromium.
Destroy a desktop VM session and free resources.
Type text on the desktop VM's virtual display.
Execute code in a sandboxed environment. Supports Python 3.12, Node.js, and Bash. Returns stdout, stderr, and exit code.
Check sandbox engine health and available tiers.
Execute a command in an isolated sandbox and return the result.
Create a persistent sandbox session with a long-lived workspace.
Destroy a session and clean up its workspace.
Execute a command in an existing persistent session.
List all sandbox sessions (active, paused, and destroyed).
Pause a session (SIGSTOP).
Read a file from a session workspace.
Resume a paused session (SIGCONT).
Write files into a session workspace.
Upload a file into a sandbox execution environment.
Desktop automation tools (desktop_click, desktop_type, desktop_key, desktop_scroll) lack coordinate validation guidance. No description of valid x,y ranges, screen dimensions, or error handling for out-of-bounds clicks.
No dependency hints or prerequisites documented. E.g., session_exec requires a valid session_id from session_create, but descriptions do not guide LLMs to call session_create first if no session exists.