Loading ToolBench
The app framework built for AI coding agents - core runtime with MCP server capabilities
Single tool 'bash' with minimal definition quality. Description is 189 chars (within baseline 194 avg), but lacks critical structure: no parameter descriptions beyond a single 'command' field, no output schema documented, no error handling guidance, no security warnings despite WRITE risk classification. Tool accepts arbitrary bash commands with no validation, constraints, or recovery guidance. This is a raw command execution interface, not a well-designed agent tool.
Run bash commands in the virtual workspace. Read files with cat, grep; navigate with ls, find; take side-effecting actions through `vobase` subcommands. The workspace layout is documented in your AGENTS.md (at `/agents/<your-id>/AGENTS.md`).
No output schema documented. Tool description does not specify what the bash command returns (stdout, stderr, exit code, structured object). LLMs cannot plan downstream operations or extract specific fields.
Parameter 'command' lacks description. Only the tool description mentions bash commands; the parameter itself has no guidance on format, constraints, or examples. LLMs cannot infer valid command syntax.
No error handling or recovery guidance. Tool marked WRITE risk but provides no guidance on what errors mean, how to retry, or what to do if a command fails. Raw bash execution with no safety rails.
Tool name 'bash' is generic and does not follow verb_noun convention. 'run_bash_command' or 'execute_command' would be clearer. Current name does not convey action.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | 2026-07-28+ | v2 |
No input validation or constraints. Tool accepts arbitrary strings with no enum, pattern, length limit, or character restrictions. LLMs can pass malicious payloads (command injection, path traversal) without guardrails.