Loading ToolBench
Open-source AI video workbench. Bring any model or your local ComfyUI, and let Claude Code or Cursor direct it over MCP. Local-first: projects, prompts, and keys stay on your machine.
Nomi exposes 24 tools with significant definition gaps. Most tools lack complete input schemas, parameters are listed but lack type definitions, constraints, and proper JSON Schema structure. Descriptions are present but often generic (8-40 chars), failing to explain WHEN to use each tool or what it returns. No output schemas are documented. Tools like 'arrange_canvas' and 'edit_timeline' use vague string parameters ('operation' type) without enums or valid value documentation. Error handling is minimal, no recovery guidance or actionable error messages. The tool set shows good domain coverage (file ops, canvas, timeline, generation) but poor LLM-readiness. Naming is mostly clear (verb_noun pattern), but parameter design forces LLMs to guess at valid inputs.
Arrange canvas nodes and connections with undo support
Ask the user a question and wait for their response. User answering is success, not failure.
Execute bash commands in the project directory with sandboxing and undo support
Cancel a running generation or export job with confirmation
Delete nodes from the canvas with confirmation
Draft video shots with parameters and optional generation start
Edit file contents in the project directory with undo support
Edit timeline with undo support
Vague operation parameters lack enums and valid value documentation. Tools like 'arrange_canvas', 'edit_timeline', 'write_script', 'read_timeline' accept a string 'operation' parameter with no list of valid values. LLMs cannot determine what operations are supported without trial-and-error.
No output schemas documented. Tools return data but LLMs cannot plan downstream calls without knowing response structure. E.g., 'look_at_canvas' has empty input schema and no documented return type.
Parameter descriptions are generic or missing type constraints. 'edit' accepts 'edits' array but does not document the structure of edit objects. 'bash' accepts 'command' string but does not warn about sandboxing limits or unsafe patterns.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 45 | 2026-07-28+ | v2 |
Export video to file with job tracking
Find files matching a pattern in the project directory
Generate video from drafted shots with approval and spend tracking
Search for patterns in files within the project directory
Read the current canvas state including nodes and connections
List directory contents in the project directory
Create an artifact node on the canvas
Read file contents from the project directory
Read document/script content
Read timeline state and structure
Save a skill to the user's library
Stage a shot on the canvas
Open the model settings panel and optionally pre-fill a provider
Undo the last timeline edit
Write file contents to the project directory with undo support
Write or modify document/script content with undo support
Destructive tools lack confirmation or dry-run support. 'delete_from_canvas' and 'cancel_job' can destroy work but offer no preview or undo mechanism beyond a generic 'undo' tool. Agents may delete unintended nodes.
No error recovery guidance. Tools do not document what errors are retryable, what requires user intervention, or what alternatives exist. E.g., 'generate' may fail due to API limits, but the error response does not suggest retry strategy or fallback.