Loading ToolBench
Start a local Multi GPT reasoning job and return immediately with a job_id. Runs usually take 5-20 minutes, commonly 10-15 minutes.
Three tools with complete input schemas and descriptions. Naming follows verb_noun pattern (multi_gpt_start, multi_gpt_status, multi_gpt_cancel). Descriptions are adequate (100-200 chars) and explain purpose and timing. However, parameter descriptions lack detail on constraints, formats, and dependencies. File attachment parameters mention 'size-limited' but no explicit max documented in schema. Model and reasoning_effort enums are well-defined but lack explanation of why they're fixed. Error handling and output schemas are not visible in provided code. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present.
Cancel a running Multi GPT background job started by this MCP server process. Terminates the active codex exec child process tree for that job and records status: canceled.
Start a local Multi GPT reasoning job and return immediately with a job_id. Runs usually take 5-20 minutes, commonly 10-15 minutes. Use this for all Multi GPT runs to avoid MCP tool-call timeouts; do not check status every minute.
Check the status of a Multi GPT background job. Returns the final result once completed. Runs usually take 5-20 minutes, commonly 10-15 minutes, so poll only occasionally unless tighter monitoring is explicitly needed.
Parameter descriptions lack constraint details. 'files' parameter mentions 'size-limited' but schema has no maxItems, and no explicit per-file or total byte limits documented in descriptions. Code shows MAX_FILE_BYTES=512KB and MAX_TOTAL_FILE_BYTES=768KB but these are not exposed to LLM via schema or description.
No output schema documented. Tools return results but LLM cannot see what fields to expect (job_id structure, status values, result format, error format). This forces LLMs to guess downstream field names and breaks tool chaining.
No tool annotations present. multi_gpt_cancel is destructive (terminates job) but lacks destructiveHint. multi_gpt_status is read-only but lacks readOnlyHint. multi_gpt_start is idempotent (same prompt+files = same job_id) but lacks idempotentHint. These hints guide agent safety reasoning.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Parameter 'model' and 'reasoning_effort' are enums with single values (gpt-5.6-luna, max) but descriptions do not explain why they are fixed or what happens if omitted. Code comment says 'Omitted values are fixed to...' but this is not in the schema description for LLM visibility.
Error handling and recovery guidance not visible in provided code. No evidence of actionable error messages, error classification (retryable vs fatal), or recovery hints. LLM cannot determine what to do if a job fails or times out.