Loading ToolBench
Pure, web-first, multi-provider TypeScript AI SDK with support for Anthropic, OpenAI, xAI, Gemini, Azure, Bedrock, and other providers. Includes agentic tool loops, streaming, structured output, and MCP integration.
This is NOT an MCP server, it is a TypeScript AI SDK (Deuz SDK) for direct LLM integration. The 4 'tools' are provider-specific integrations (Anthropic/OpenAI/Google web search, plus a sub-agent tool), not MCP tool definitions. Tool schemas are partially visible but lack critical structure: googleSearch has an empty input schema ({}), agentTool has minimal schema, and descriptions are generic. No evidence of MCP protocol implementation, transport, or server-mode operation. This evaluation treats the 4 items as if they were MCP tools for scoring purposes, but the fundamental mismatch (SDK vs. MCP server) explains the low score.
Turns an agent into a tool the parent loop can call (a sub-agent). Unlike a black-box delegate, a sub-agent's whole canonical stream is forwarded LIVE into the parent when the parent is streaming, and its tool calls stay gated by the parent's server-mode approveToolCall.
Anthropic server-side web search tool that runs on the provider during a turn. The model decides to call it, the provider executes it server-side, and results stream back inside the same response.
Gemini Google Search grounding tool that runs on the provider during a turn. The model decides to call it, the provider executes it server-side, and results stream back inside the same response.
OpenAI Responses hosted web search tool that runs on the provider during a turn. The model decides to call it, the provider executes it server-side, and results stream back inside the same response.
googleSearch has empty input schema ({}). No parameters documented. LLM cannot determine what inputs are valid or required.
Tool descriptions are generic and lack actionable context. 'Gemini Google Search grounding tool' does not explain WHEN to use it vs. anthropicWebSearch or openaiWebSearch, or what the LLM should expect in the response.
agentTool schema is minimal (only 'prompt' required, no output schema documented). LLM cannot plan downstream tool calls or understand what a sub-agent returns.
No error handling guidance. If a web search fails or a sub-agent times out, the LLM has no recovery path documented.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 32 | 2025-06-18+ | v2 |
This is an SDK, not an MCP server. No evidence of MCP protocol transport, tool registration via MCP, or server-mode operation. Evaluation assumes these are MCP tools, but the codebase is a direct LLM integration library.