Loading ToolBench
JSON-RPC (MCP) over HTTP for one tenant's archive. Provides access to recordings, transcripts, summaries, and semantic search over a tenant-private SQLite archive.
Five tools with complete JSON Schema input definitions and clear descriptions. Naming follows verb_noun pattern (recordings_list, recordings_search, recording_get). Descriptions are concise (34 - 95 chars) and explain purpose well. Schemas include proper type constraints, enums where appropriate (e.g., anyOf for recording number formats), and pagination support (limit, cursor). However, output schemas are not documented, LLMs cannot predict response structure. Parameter descriptions lack detail on expected formats and constraints (e.g., what does 'opaque cursor' mean to an LLM?). No error handling guidance visible. No tool annotations (readOnlyHint present in risk field but not in MCP schema). Missing descriptions for nested filter object properties.
One recording: metadata, the preferred transcript (ASR, falling back to PLAUD), the generated Russian summary, and which ASR engine produced it.
Structured mind-map data for one recording.
Hybrid lexical and tenant-local semantic search over this archive. Exact lexical matches rank first.
List this archive's recordings, newest first, one bounded page at a time.
Full-text search (FTS5) over recording names and stored transcripts.
Output schemas not documented. LLMs cannot predict response structure (fields, types, nested objects). Responses for recording_get include 'metadata, transcript, summary, ASR engine' but schema is invisible to the agent.
Nested object properties lack descriptions. 'filters' object in recordings_list has date_from, date_to, min_duration_ms but no descriptions for these fields. LLMs cannot infer what 'min_duration_ms' means or its valid range.
Parameter descriptions lack format/constraint details. 'cursor' described as 'Opaque cursor from next_cursor', what format? How long? Can it be reused? LLMs need explicit constraints to avoid malformed requests.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
No error handling guidance. Tools return READ_ONLY risk but no documented error responses (e.g., 'recording not found', 'invalid cursor', 'search timeout'). LLMs cannot plan recovery.
Tool annotations missing from MCP schema. Risk field (READ_ONLY) is metadata but not exposed as readOnlyHint in the MCP tool definition. Agents cannot infer safety from the schema.