Loading ToolBench
A governed control plane for MCP-connected AI agents and tools
AgentGrid exposes 16 control-plane and approval-management tools with basic schemas and descriptions. All tools have names starting with action verbs (create_, update_, delete_, list_, get_, discover_, request_, decide_) and descriptions present. However, most descriptions are terse (10 - 50 chars), parameter descriptions are minimal or absent, and output schemas are not documented. Input schemas are present but lack depth: many parameters have no description, type constraints are sparse, and error handling guidance is missing. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. The server is STDIO-only, capping protocol readiness at 50 regardless of definition quality.
Registers a new MCP server with the control plane
Creates a control-plane policy to govern tool execution
Registers a tool discovered from an MCP server
Makes an approval decision (approve or reject) for an execution awaiting approval
Deletes an MCP server and its associated tools and policies
Deletes a control-plane policy
Discovers tools available from an MCP server by connecting to it and retrieving its tool list
Output schemas not documented. LLMs cannot infer what fields list_mcp_servers, get_mcp_server, list_tools, list_policies, get_policy, get_approval return. This forces agents to guess structure and risks failed downstream tool chaining.
Parameter descriptions missing or trivial. 'server_id' appears in 7 tools with identical description 'The ID of the server...' but no guidance on format, range, or how to obtain it. 'configuration' in create_mcp_server is described as 'Server configuration including command, args, env, cwd, timeout_ms' but lacks structure, required fields, or examples.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
Retrieves approval status for an execution
Retrieves details of a specific MCP server
Retrieves details of a specific control-plane policy
Lists all registered MCP servers
Lists all control-plane policies
Lists all tools for a specific MCP server
Requests human approval for a pending execution
Updates an existing MCP server configuration
Updates an existing control-plane policy
No tool annotations despite clear risk levels. delete_mcp_server, delete_policy, and decide_approval are destructive/irreversible but lack destructiveHint annotation. create_mcp_server, update_mcp_server, create_policy, update_policy lack idempotentHint guidance. discover_mcp_server_tools, list_mcp_servers, get_mcp_server, list_tools, list_policies, get_policy, get_approval lack readOnlyHint.
No error handling guidance. Tools lack descriptions of failure modes, recovery steps, or actionable error messages. E.g., create_mcp_server does not document what happens if 'command' is invalid, if 'timeout_ms' exceeds limits, or how to retry on transient failures.
Pagination not documented. list_mcp_servers, list_tools, list_policies return lists but lack page/offset/limit parameters or total_count in response. Large result sets will blow context windows without pagination guidance.