Loading ToolBench
Unofficial GMI Cloud CLI (+ MCP server) for GMI Studio — image/video/audio generation and LLM inference
Server has 7 tools with clear verb-noun naming (list_, get_, generate_, upload_, chat). All tools have descriptions (avg 120 chars, within baseline 34-392). Input schemas are present and typed with Zod. However, parameter descriptions are sparse or missing: 'temperature' and 'max_tokens' in chat tool have null descriptions (critical gap). Output schemas are undocumented, LLMs cannot infer what fields to expect from generate_media, get_generation, or list_llm_models. Error handling returns GmiError messages but lacks recovery guidance (e.g., 'retry with wait=false' for timeouts). No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. generate_media and upload_file are destructive but lack confirmation patterns.
Run a chat completion against a GMI Cloud LLM (e.g. DeepSeek, Qwen, Llama). Returns the assistant's reply text.
Submit an image/video/audio generation job to GMI Cloud Studio. By default waits for completion and returns media URLs. Set wait=false to get a request_id back immediately (use get_generation to poll).
Check the status of a generation job. When status is 'success', media URLs are included.
Get a Studio model's parameter schema, defaults, and pricing. Call this before generate_media to build a valid payload.
List GMI Cloud Inference Engine LLM models (OpenAI-compatible chat API).
List GMI Cloud Studio generative-media models (image, video, audio generation). Use get_studio_model to see a model's parameters before generating.
chat tool has two parameters (temperature, max_tokens) with null descriptions, violates pattern:tool-description requirement that every parameter must have a non-empty description.
No output schemas documented for any tool. LLMs cannot infer what fields generate_media, get_generation, list_studio_models, or list_llm_models return, forcing them to guess downstream field names and risking broken tool chains.
Destructive tools (generate_media, upload_file) lack confirmation patterns or dry-run support. Agents cannot preview results before spending credits or uploading files.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Upload a local file (e.g. a reference image for image-to-video) to GMI Cloud and get back a public URL usable in generate_media payloads.
Error messages lack recovery guidance. GmiError responses return raw error text without actionable next steps (e.g., 'retry with wait=false', 'call list_studio_models first').
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels declared in the spec. generate_media and upload_file are WRITE; others are READ_ONLY.