Loading ToolBench
An Express-based server that processes context through an LLM (Mixtral via OpenRouter) for agent workflows
Single tool 'mcp/context' has a reasonable input schema with typed parameters and descriptions, but critical gaps undermine quality. Tool name violates verb-first convention (should be 'process_context' or 'send_to_llm'). Description is adequate (89 chars) but lacks clarity on when/why to use it vs alternatives. Output schema is completely undocumented, callers cannot know what fields to expect from the response. No error handling guidance, no pagination support, no idempotency markers. The tool conflates two concerns: LLM processing AND context enrichment. No security considerations for API key injection or rate limiting.
Processes context input through an LLM (Mixtral) and returns an updated context with the LLM response, agent ID, and timestamp
Tool name 'mcp/context' does not start with action verb. Should be 'process_context', 'send_to_llm', or 'enrich_context' to signal intent.
Output schema completely undocumented. Response structure (updatedContext fields, success flag, error format) is inferred from code, not declared. LLMs cannot plan downstream calls without knowing return fields.
No error handling guidance. Errors return raw HTTP 500 with message only. LLM cannot determine if error is retryable, user-fixable, or fatal. No recovery suggestions.
Tool combines two responsibilities: calling external LLM AND enriching context object. Should split into separate tools (call_llm + enrich_context) for composability.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 28 | 2026-07-28+ | v2 |
No rate limiting, timeout handling, or API key validation. OpenRouter API key exposed in environment; no per-request secret injection pattern. Tool can be abused to exhaust quota.