Loading ToolBench
MCP server that executes Claude Code commands in a project directory and integrates with Slack for command execution and result reporting
Single tool 'claude_code' has critical gaps across all dimensions. The tool definition is visible in claude-code-mcp/index.js (lines 28-48 in the excerpt) but lacks depth and rigor expected for production MCP tools. The schema is present but minimal; descriptions are present but generic; no error handling guidance; no input validation examples; tool name lacks clear verb-noun pattern for the action being performed. The cwd parameter is optional but no guidance on default behavior. The prompt parameter is required but no constraints, examples, or failure modes documented. Output schema is completely undocumented, LLMs cannot know what fields to expect from successful execution. No mention of what 'execute Claude Code commands' means operationally. Given MCP baseline from 549 production tools, this falls well below community standards.
Execute Claude Code commands in a project directory
Tool name 'claude_code' lacks action verb. Should be 'execute_claude_code' or similar verb_noun pattern to clarify intent for LLMs.
No output schema documented. LLMs cannot infer what fields the tool returns (error messages, results, status, logs, file changes, etc.). This prevents downstream tool chaining and forces the agent to guess at response structure.
Tool description 'Execute Claude Code commands in a project directory' is vague. What does 'execute' mean? Run tests? Generate code? Compile? What commands are valid? When should this be called? No context for LLM selection.
Parameter 'prompt' lacks constraints or format guidance. No description of what Claude Code 'commands' or 'requests' are valid. LLMs will hallucinate arbitrary prompts. Should document expected format, valid command types, or reference Claude Code CLI documentation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 35 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 27 | - | v1 |
Parameter 'cwd' defaults to PROJECT_PATH but this is vague. What if PROJECT_PATH is undefined? Does the tool fail silently? What are the side effects of operating in different directories? No guidance on validation or error cases.
No error handling documentation. What happens if Claude CLI is not installed? If the cwd is invalid? If the prompt triggers a crash? No recovery guidance for LLMs. Bare error messages from spawn() will not be actionable.
No idempotence contract. If the prompt is 'create a file', is it safe to retry? Will the tool error if the file already exists, or silently overwrite? Agents need to know which calls are safe to retry.