Loading ToolBench
A web-based dashboard for managing Claude extensions, profiles, and code analysis. Provides inventory scanning, dependency graphing, bisect debugging, import/export, and MCP server tool management.
This HTTP-based Next.js dashboard exposes 17 tools with significant definition gaps. Most tools lack input parameter descriptions entirely (tools 1, 3, 6, 10 have no visible schemas). Descriptions are present but generic (avg ~80 chars, below the 194-char baseline). No tool declares output schemas, error recovery guidance, or permission requirements. Tool naming is HTTP-verb-based (GET/POST/DELETE/PATCH) rather than action-verb-based (get_*, create_*, delete_*), which is non-standard for MCP. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. Parameter descriptions are minimal or absent across all tools.
Delete a single dashboard item with rate limiting and path validation
Get current bisect session state and status
Get top-N entries by total cost for a given dimension (agent, skill, or cwd)
Build and return the dependency graph for all scanned items
Liveness probe endpoint that returns server uptime and health status
Scan and return all dashboard items with optional filtering by category and scope
Check JSONL to SQLite migration progress
List all available settings profiles
Tool names use HTTP verbs (GET, POST, DELETE, PATCH) instead of action verbs (get_, create_, delete_, update_). LLMs expect verb_noun naming to infer intent from the name alone.
8 tools (health, dependencies, stats, migration-status, bisect GET, profiles GET) have no visible input schemas. Cannot verify parameter types or constraints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 42 | 2026-07-28+ | v2 |
Load and parse a session JSONL file, extracting user and assistant messages with token counts
Full-text search across user message previews via SQLite FTS5
Retrieve daily activity statistics, weekly top 5 items, tool usage, and daily costs from SQLite database
Toggle an individual MCP tool's enabled/disabled state in settings
Perform bisect actions: start a new session, provide feedback on reproduction, or abort current session
Export dashboard items as a profile in 'all' or 'selected' mode
Apply an import profile with merge, overwrite, or clean mode
Preview what an import profile would change without applying it
Manage settings profiles: save, switch, or delete a profile
No tool declares output schemas. LLMs cannot plan downstream calls or extract required fields (e.g., IDs for chaining). Baseline: 100% of A+ tools document return types.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels (DELETE, POST, PATCH). Agents cannot determine which calls are safe to retry.
Parameter descriptions are minimal or absent. E.g., 'category' and 'scope' in /api/inventory lack detail on valid values, format, or constraints. Baseline: 100% of A+ tools have param descriptions.