Loading ToolBench
MCP server that delegates Claude Code subagents to alternative backends (local llama.cpp, LiteLLM, DeepSeek, Bedrock, etc.) with full tool calling.
The 'delegate' tool has a well-structured input schema with 5 parameters, each with type and description. The tool description is substantive (165 chars) and explains the delegation mechanism. However, there are notable gaps: (1) no output schema is documented, callers cannot know what the tool returns or how to chain results; (2) parameter descriptions lack detail on format/constraints (e.g., what constitutes a valid agent name, expected model format, semantics of max_turns); (3) no explicit error handling documentation or recovery guidance; (4) the tool performs multiple responsibilities (agent lookup, LLM backend delegation, tool calling orchestration), approaching a composition concern. The schema is present and typed, placing it above the median, but incomplete documentation and lack of output schema definition prevent a higher score.
Delegates task execution to an alternative LLM backend (LiteLLM, llama.cpp, DeepSeek, Bedrock, etc.) with agentic tool calling, preserving the Claude Code orchestrator's plan and session.
No output schema documented. Tool returns results from agentic loop but callers cannot see what fields to expect, breaking tool chaining and forcing LLMs to guess structure.
Parameter descriptions lack actionable constraints. 'agent' param says 'Agent name from ~/.claude/agents/*.md' but does not specify file format, required fields in .md, or error behavior if agent not found. 'model' says 'Optional model alias' without explaining alias format or valid values. 'max_turns' lacks range/bounds.
No error handling or recovery guidance documented. Tool can fail on agent lookup, backend connectivity, token exhaustion, timeout. Tool description does not indicate what errors are retryable vs fatal, or how LLM should respond.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
Tool combines multiple responsibilities: agent configuration lookup, backend LLM invocation, agentic loop orchestration, tool calling (read_file/write_file/run_bash), and session preservation. Violates single-responsibility principle, should split into separate tools or explicitly document the unified concern.
Parameter naming inconsistency: 'max_turns' implies iteration count, but tool description references 'LOCAL_MAX_TURNS' and 'CLOUD_MAX_TURNS' env vars, suggesting backend-specific logic. Parameter description does not clarify how per-model defaults apply if parameter omitted.