WebMCP Toolbench

Inspect, run and test WebMCP tools, manually or with the AI provider of your choice.

ebMCP Toolbench is a side panel for developers building WebMCP tools: the tools a web page registers so that AI agents can use it. Open it on a page to see every tool, run them by hand, and let an AI agent use them with the model and provider of your choice. REQUIREMENTS • Chrome 155 or later • WebMCP turned on: open chrome://flags/#enable-webmcp-testing, set it to Enabled and relaunch Chrome • An API key only if you want to use the agent (OpenAI, Anthropic, Google Gemini, OpenRouter or Groq), or a local model (Ollama, LM Studio), or Chrome's built-in AI where your computer supports it WHAT IT DOES • Inspect: the tools in the page and in every iframe, with their input schemas, annotations, origin, and whether the top-level page can see them. Form-based (declarative) tools show each field next to the parameter Chrome made of it. • Lint: WebMCP rules, description quality, what changes when a tool is sent to your provider, and instruction-like or hidden text in descriptions. • Run: from a form generated from the schema or from raw JSON. Arguments are validated, runs can be cancelled or time out, and every run is kept in a history, including results that arrive through a navigation or a new tab. • Agent: chat with an AI model that calls the page's tools. Tools that may change something ask first, and you can edit the arguments. • Voice (optional, off by default): talk to the agent through Gemini Live, hands-free or push-to-talk. • Override: try another title, description or schema for a tool as the agent sees it, without redeploying. The page is never changed. • Changes: how the page's tools change over time, with field-level diffs. • Logs and model calls: a filterable timeline of events, and every request sent to the model with its streamed response, like DevTools' Network panel. • Session bundles, replay and evals: save a tab's tools, runs, conversation and logs to one file, replay a conversation with another model, and export webmcp-evals test cases. • Export: tools as JSON, an MCP tools/list result, OpenAI, Anthropic or Gemini function declarations, registerTool() code or ScriptToolConfig. PRIVACY No account, no analytics, and no servers of ours. Toolbench runs in a page only while its panel is open on that tab. If you use the agent, your messages, the page's tools and their results go to the AI provider you choose, and only there. API keys stay in your browser. Privacy policy: https://webmcp-toolbench.solvemybizz.org/privacy Try it on the demo page: https://webmcp-toolbench.solvemybizz.org/demo/ WebMCP Toolbench is an independent project and is not affiliated with or endorsed by Google.