Loading ToolBench
Turn any laptop, desktop, VM, or server into a secure AI-agent workspace via MCP. Let ChatGPT, Claude, Grok, Qwen and other MCP clients read and edit files, run commands, test code, inspect systems, and complete real-world tasks.
Only 2 tools visible in source code (capability_discover, capability_batch). Both have descriptions and input schemas, but schemas lack critical detail: no parameter types defined, no required field specifications, no output schemas documented. capability_discover description is 127 chars (acceptable range), but parameter descriptions are minimal. capability_batch description is 135 chars but similarly sparse on parameter detail. The server registers 30+ tool suites (system, filesystem, shell, git, sqlite, etc.) but actual tool definitions are not visible in provided source, only registration calls. This server fails the evidence-based evaluation criterion.
Run multiple independent read-only tool calls in parallel. Maximum 8 calls; mutating tools are rejected.
Discover only the capabilities available to the current token. Use query/category/cost to keep context small.
Input schemas lack type definitions for parameters. capability_discover and capability_batch show parameter names and descriptions but no 'type' field (string, number, boolean, array). JSON Schema requires explicit types; LLMs cannot infer parameter types from descriptions alone.
No output schemas documented for any visible tool. capability_discover and capability_batch do not specify what fields are returned, their types, or structure. LLMs cannot plan downstream tool calls without knowing response shape.
Tool definitions inferred from registration calls, not directly visible in source. src/tools/index.ts shows 30+ registerXxxTools() calls but actual tool definitions (names, schemas, descriptions) are in separate files not provided. Cannot verify schema completeness, parameter validation, or error handling for 30+ tools.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
capability_batch parameter 'calls' is an array of objects with 'name' (required) and 'arguments' (optional), but 'arguments' has no schema, it is a free-form object. LLMs cannot validate argument structure before submission; any malformed arguments will fail at runtime.
No error handling guidance in visible tool descriptions. capability_discover and capability_batch do not document what errors can occur, how to recover, or what the LLM should do if a call fails (retry, ask user, abort).