Loading ToolBench
Cross-platform desktop shell for agent CLIs and ACP-speaking agents. Provides host MCP tools for filesystem operations, process management, networking, and agent spawning.
ShellX exposes 11 tools with inconsistent quality. Most tools have basic descriptions (10-50 chars), but lack depth on WHEN to use them or what they return. Input schemas are present but sparse: parameters lack type constraints (no enums for action strings), and output schemas are entirely undocumented. Tool names are verb-noun style (good), but descriptions are too terse to guide LLM selection. No error handling guidance, no pagination support, no idempotency hints. The Agent tool is particularly concerning, it spawns subagents with multiple timeout parameters (wait_budget_ms, max_runtime_ms, timeout_ms) without clarifying which takes precedence or what happens on timeout. Security: no evidence of secret injection or permission gates. Composition: tools are narrowly scoped (good), but lack chaining IDs in responses (e.g., build_state returns tabId but unclear if it chains to other tools).
Spawn a subagent with a specified persona to execute a task
Create a git checkpoint for the current build session
Query the current state of a build session
Perform actions on ShellX Cut tooling (open, check status)
Read ShellX Cut tooling status and metadata
Watch a filesystem path for changes with optional debouncing
Run diagnostics on a work preview session
Retrieve logs from a work preview session
Output schemas completely undocumented. No tool specifies what fields it returns, forcing LLMs to guess structure and breaking downstream tool chaining.
Action parameters (cut_read, cut_act) accept free-form strings without enum constraints. LLMs will hallucinate invalid actions like 'status_check' or 'open_editor' instead of exact values.
Agent tool has three conflicting timeout parameters (wait_budget_ms, max_runtime_ms, timeout_ms). Descriptions do not clarify precedence, defaults, or behavior on timeout. LLMs cannot reason about which to use.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2025-06-18+ | v2 |
Start a work preview session
Get the current state of a work preview session
Stop a work preview session
Descriptions are too terse (10-50 chars). They state WHAT the tool does but not WHEN to use it, what it returns, or how it chains with other tools. Baseline for A+ tools is 50-200 chars with context.
No error handling guidance. Tools do not document what errors are retryable, what user-fixable errors look like, or how to recover. LLMs will not know whether to retry or escalate.