Loading ToolBench
Audio intelligence for agents. An MCP server that lets agents download, transcribe, search, and analyze audio.
Augent provides 20 tools with complete input schemas and descriptions for all tools and parameters. Naming follows verb_noun convention (transcribe_audio, search_audio, export_clips). However, descriptions are often generic (e.g., 'Transcribe an audio file using faster-whisper with word-level timestamps' is 80 chars, below the 194-char baseline). Output schemas are not documented, LLMs cannot predict return structure. Error handling is absent from descriptions. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Parameter descriptions lack format constraints (e.g., 'model_size' lists values but not as enum). Tools like search_audio, search_audio_full, and search_audio_proximity have overlapping functionality without clear disambiguation. Composition is weak: three separate search variants (search_audio, search_audio_full, search_audio_proximity) should be one configurable tool.
Clear transcription memory cache
Clear loaded model cache
Semantic search in audio by meaning using sentence embeddings
Auto-detect topic boundaries and chapters in audio
Extract audio clips around keyword matches
Export keyword matches to various formats (CSV, SRT, VTT, Markdown)
Export transcription to SRT or VTT subtitle format
Output schemas not documented. LLMs cannot predict return structure for any tool. Descriptions state what tools do but not what they return (fields, types, pagination). This violates pattern:tool and forces agents to guess downstream field names.
Three overlapping search tools (search_audio, search_audio_full, search_audio_proximity) lack clear disambiguation. LLMs cannot determine which to call without trial-and-error. Violates pattern:tool composition, should be one configurable search_audio with optional parameters (include_metadata, proximity_keyword, max_distance).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Get cached transcription by title
Get transcription memory statistics
Extract vocal stem from audio
Identify and diarize speakers in audio (requires pyannote-audio)
List all cached transcriptions
Read transcription or text aloud using text-to-speech
Render text as ASCII art banner with color support
Transcribe audio and search for keywords with timestamps
Full output including transcription metadata and keyword matches
Find where keyword1 appears near keyword2 in audio
Separate audio into vocal and instrumental stems using Demucs
Convert text to speech using Kokoro TTS
Transcribe an audio file using faster-whisper with word-level timestamps
Parameter descriptions lack format constraints. 'model_size' lists values ('tiny, base, small, medium, large') as prose instead of enum constraint. 'device' and 'compute_type' similarly lack formal enums. LLMs cannot validate inputs and may hallucinate invalid values.
Tool descriptions are generic and under 100 chars (baseline 194 chars). E.g., 'Get transcription memory statistics' (40 chars) does not explain when to call it, what it returns, or why. Descriptions should answer: What? When? Why? What does it return?
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications in metadata (READ_ONLY vs WRITE). Agents cannot determine which tools are safe to retry or have side effects. Violates current MCP spec (2026-07-28) tool annotations pattern.