Loading ToolBench
AI Agent I/O Tracking & Rollback System — MCP Proxy Gateway
AgentsGate provides 14 tools across database and filesystem domains with explicit schemas and descriptions. Most tools have clear verb-noun naming (query, execute, read_file, write_file) and documented parameters with types. However, descriptions are often terse (10-50 chars), parameter descriptions lack actionable constraints, and output schemas are not documented. Error handling is minimal, no recovery guidance or categorization. The database tools (execute, execute_ddl) lack confirmation/dry-run patterns for destructive operations despite high risk. Test helper tools (echo, fail, slow, inspect_request) are present but not production-grade. Average tool description length ~45 chars (well below 194-char baseline); most params lack format/range guidance.
Delete a file by moving it to a .trash folder instead of permanently deleting it
Delete a table snapshot
Echoes the message argument back as a text content item.
Edit a file by specifying old text to replace with new text
Execute a DML statement (INSERT, UPDATE, DELETE)
Execute a DDL statement (CREATE, ALTER, DROP, TRUNCATE)
Always returns a JSON-RPC application error (code -32000).
Destructive operations (execute_ddl with DROP/TRUNCATE) lack dry-run or confirmation pattern. confirm_destructive boolean exists but no guidance on recovery if agent makes mistake.
Tool descriptions are terse (avg 45 chars vs 194-char baseline). Most lack WHEN to use, prerequisites, or state-modification warnings. E.g., 'Execute a DML statement' does not explain idempotency or retry safety.
Parameter descriptions lack actionable constraints. 'sql' param has no format guidance; 'limit' has min/max but no explanation of default behavior when omitted. 'path' param has no path traversal or symlink guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Returns the raw tools/call params as JSON text — useful for verifying that proxy middleware (e.g. _agentsgate) was injected.
List files and directories in a directory
List all table snapshots
Execute a read-only SQL SELECT statement
Read the contents of a file
Waits `delay` milliseconds before responding.
Create a new file with the given contents
Output schemas not documented. Tools return results but LLM cannot plan downstream calls without knowing field names, types, and structure. E.g., query() returns rows but schema is implicit.
Error handling provides no recovery guidance. No categorization of retryable vs user-fixable vs fatal errors. No suggestions for alternative tools or next steps on failure.