Loading ToolBench
MCP server that gives an AI assistant eyes for video, via the vidgrid render API.
Two tools with reasonable descriptions and basic schemas. render_video has a detailed, LLM-optimized description (180+ chars) explaining WHAT, WHEN, and HOW to use it. check_balance is minimal but functional. Both tools have input schemas with types and descriptions. However, output schemas are not documented, LLMs cannot predict what fields to expect from responses. Parameters lack constraints (grid enum, transcribe default behavior). No error handling guidance. No tool annotations (readOnlyHint). Composition is sound (single responsibility), but schema completeness and error recovery patterns are missing.
Return the remaining Vidgrid render credits for the configured API key.
Turn a local video FILE into numbered frame-grid images (+ transcript) you can read. Use this whenever you need to "watch" or analyze a video you can't otherwise see: summarize a talk, find the exact moment something happens, read on-screen text/UI, or rank clips. Returns one or more grid PNGs — cells are numbered globally and each is tagged with its timestamp — plus the spoken transcript. Reason over the frames + transcript to answer; reference any frame by its number.
Output schemas not documented. LLMs cannot predict response structure (grid PNGs, transcript format, credits_remaining field). Forces agents to guess at downstream field names.
No error handling guidance. Tool descriptions lack recovery hints (e.g., 'If file is too large, try a shorter clip' or 'If API key is invalid, check VIDGRID_API_KEY env var'). Agents cannot self-correct on failures.
grid parameter lacks enum constraint. Description says '2x2|3x3|4x4|5x5' but no formal enum in schema, LLMs may pass invalid values like '6x6' or 'auto'.
check_balance description is 50 chars, below the 10 - 1024 baseline. Lacks context: when should the agent call this? What does 'credits' mean? Is it per-request or account-wide?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
No tool annotations. Both tools are read-only (no state mutation), but readOnlyHint is not set. Agents cannot infer safety/idempotency without explicit hints.