Loading ToolBench
MCP server for integrating Blender with Claude via Model Context Protocol, enabling 3D model generation and manipulation through natural language prompts
This server provides 6 tools for MCP server orchestration (discovery, search, connection, execution). While tool names follow verb-noun conventions and descriptions are present, there are critical gaps in schema completeness, parameter documentation, and error handling. Most tools have minimal input schemas with little constraint detail, and output schemas are not formally documented. The semantic search tool's async-only implementation and missing pagination support for list results are concerning for production use.
Connect to a specific MCP (Model Context Protocol) server to use its tools. Use this tool to connect to a specific server and use its tools.
Disconnect from the currently active MCP (Model Context Protocol) server
Get the currently active MCP (Model Context Protocol) server
Lists all available MCP (Model Context Protocol) servers that can be connected to, along with the tools available on each server. Use this tool to discover servers and see what functionalities they offer.
Search for relevant tools across all MCP servers using semantic search. Provide a description of the tool you think you might need to be able to perform the task you are assigned. Do not be too specific, the search will give you many options. It is important you search for the tool, not for the goal. If your first search doesn't yield relevant results, try using different keywords or more general terms.
Execute a specific tool on a specific server without first connecting to it. This is a direct execution shortcut that combines connection and tool execution into a single step. Specify the server name, tool name, and the input to the tool.
No output schemas documented for any tool. LLMs cannot infer return types, field names, or structure. This violates the critical pattern that 100% of A+ tools have documented return types.
search_mcp_tools accepts 'top_k' with default=100 but no maximum constraint documented. Returning 100+ results in a single response risks context window exhaustion. Baseline from rubric: 'cap results at reasonable limit (20-50) and offer pagination.'
use_tool_from_server parameter 'tool_input' is overloaded as 'object|string'. This violates the pattern of explicit typed parameters and forces the LLM to guess the correct format. Should be split into separate tools or clearly constrained.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 9 | <=2025-11-25 | v2 |
| 2026-03-09 | D | 52 | - | v1 |
search_mcp_tools is async-only (raises NotImplementedError in _run). This limits integration with synchronous MCP clients and breaks the LangChain tool protocol contract. All tools should support both sync and async.
No error handling guidance documented. When a tool fails (e.g. 'server not found'), there is no documented error response format, classification (retryable vs fatal), or recovery guidance. Violates pattern:recovery-guide.
Descriptions for get_active_mcp_server and disconnect_from_mcp_server are under 70 characters and lack sufficient context. Baseline: descriptions should be 10-1024 chars with WHAT, WHEN, and RESULT details.
No parameter constraints for server_name or tool_name. These accept free-form strings but should either be enums (list of available servers) or have format patterns documented.
No idempotency or confirmation patterns documented. use_tool_from_server is marked WRITE but offers no dry-run or confirmation step. Agents making mistakes could trigger unintended operations on remote servers.