Loading ToolBench
Expose web-only GPT Pro (and any ChatGPT model) to MCP clients via your real, logged-in browser
Single tool 'ask_gpt_pro' has a well-written description (194 chars, within baseline 34-392 range) explaining what it does, when to use it, and what it returns. Input schema is present with two parameters: 'prompt' (required, string) and 'conversation_id' (optional, string). Both parameters have descriptions. However, the schema lacks formal constraints (no enums, patterns, or min/max bounds). Output schema is documented in the description ('Returns {text, model, conversation_id}') but not formally specified in the tool registration. No error handling guidance is provided, the tool description does not explain what happens on failure or how to recover. The tool name 'ask_gpt_pro' follows verb_noun convention and is clear, but the implementation is opaque (ChatGPTDriver internals not shown), making it difficult to assess idempotency, timeout handling, or retry safety.
Ask GPT Pro (ChatGPT web, Pro model) and return its full reply. GPT Pro reasons for several minutes; this call BLOCKS until the answer is complete. The prompt must be self-contained (Pro shares no other context). Pass `conversation_id` from a previous reply to continue the same thread — essential for a multi-turn debate so Pro remembers the earlier rounds. Returns {text, model, conversation_id}.
Output schema not formally declared. Description states 'Returns {text, model, conversation_id}' but no JSON Schema output type is registered. LLMs cannot reliably extract fields or plan downstream operations.
No error handling guidance. Description does not explain failure modes (network timeout, GPT Pro unavailable, invalid conversation_id) or recovery steps. Agents cannot self-correct on failure.
Input parameters lack formal constraints. 'prompt' has no length bounds; 'conversation_id' has no format validation. Free-form strings invite invalid input. Should specify: prompt length (e.g. 1-10000 chars), conversation_id format (UUID pattern).
Blocking behavior not clearly signaled. Description mentions 'this call BLOCKS until the answer is complete' and 'reasons for several minutes', but no timeout parameter or guidance on expected duration. Agents may not know to expect long waits.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | <=2025-11-25 | v2 |
No idempotency guarantee documented. Unclear whether retrying the same prompt+conversation_id is safe or will create duplicate messages. Critical for agent retry logic.