Loading ToolBench
MCP server for recording human-in-the-loop workflows and generating skill files with VideoDB capture and compilation
VideoDB Record & Replay has 5 tools with basic schemas and descriptions, but significant gaps prevent production use. Tool names are action-oriented (request_*, record_*, stop_*, compile_*, list_*), which is good. However, descriptions are inconsistent in depth, some are detailed (record_skill_tool: 280+ chars explaining human-in-the-loop workflow), others are terse (list_skills_tool: 50 chars). Input schemas are present but minimal: most tools have 0-2 parameters with basic type info. Critically, NO output schemas are documented anywhere in the visible code, LLMs cannot predict what these tools return, breaking downstream composition. Error handling is absent: no guidance on what happens if recording fails, compilation errors, or permissions are denied. The tool descriptions lack dependency hints (e.g., 'Call request_capture_permissions_tool before record_skill_tool'). Parameter descriptions are sparse, 'lead_in_seconds' and 'trim_end_seconds' lack range constraints or examples. No idempotency guarantees, no retry guidance, no recovery paths.
Compile a recording into SKILL.json and SKILL.md files. After creating SKILL.md, install it into the agent's global skills directory so the generated skill is available for future agent runs.
List all skills generated through this MCP.
Start a human-in-the-loop workflow recording. The agent should call this tool, tell the human operator when recording is active and when to begin after the lead-in, then wait. The human performs the UI actions being recorded. Do not drive the target app with browser, desktop, or shell automation during recording unless the user explicitly asks the agent to demonstrate the workflow.
Ask the system for VideoDB capture permissions before recording.
Stop the active human-operated recording after the user says stop.
No output schemas documented. LLMs cannot predict return types, breaking tool composition and forcing agents to guess what fields are available for downstream calls.
Parameter descriptions lack constraints. 'lead_in_seconds' and 'trim_end_seconds' have no range (min/max), units, or examples. LLMs may pass invalid values (negative, absurdly large).
No error handling or recovery guidance. If recording fails, permissions denied, or compilation errors occur, LLMs have no actionable next steps.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Missing tool dependencies and sequencing hints. No description states 'Call request_capture_permissions_tool before record_skill_tool' or 'compile_skill_tool requires a valid video_id from a prior recording'.
list_skills_tool description is vague (50 chars). Does it return skill names, full metadata, file paths? LLMs cannot determine what to do with the result.