WebMCP Explorer

Inspect, test, debug and review the security of the WebMCP tools a website exposes to AI agents.

WebMCP Explorer is a developer workbench for inspecting, testing, debugging and reviewing the tools a website exposes to AI agents through WebMCP. Discover tools, inspect their schemas, run them manually and review their security — no AI account or API key required. INSPECT WHAT THE PAGE EXPOSES See tool names, descriptions, input schemas and declared annotations. Identify the frame each tool belongs to, including accessible iframes. Explorer reports incomplete discovery when a frame cannot be inspected. Follow registrations, removals and executions in the timeline, and inspect contract-quality findings in Audit. Audit findings grade the contract, not safety; the security review is the Security section below. RUN TOOLS AND KEEP REUSABLE CHECKS Fill a generated form or enter raw JSON, then read the tool's result and execution timing. Cancel a slow call. Save argument sets and expected outcomes, including checks on selected JSON fields. When a tool or schema changes, Explorer flags affected saved checks for review. BUILD MANUAL RECIPES Combine reviewed calls into a sequence with named inputs and explicit use of earlier results. Run step by step, pause or stop, inspect each outcome and export a report. Recipes work without an AI model. Interrupted actions are not automatically repeated. REPRODUCE AND COMPARE Export tool contracts, results and diagnostics as JSON or Markdown. Turn a recorded call into a DevTools console snippet or a standalone Playwright script. Save contract snapshots and compare changes between versions or environments. Preview and edit diagnostic evidence before sharing it. TEST LOCAL CHANGES Try alternative descriptions or schemas and simulate responses such as errors, timeouts or malformed JSON. Overrides are clearly marked, and the original contract remains visible. Use eval suites with your selected AI provider to compare which tools a model chooses before and after a change. OPTIONAL: CONNECT YOUR OWN AI Connect a supported cloud provider with your own API key, or use a compatible local model server. No provider is selected automatically. Cloud providers use your own API key and may charge for usage; local servers require a running model and suitable hardware. Check the endpoint connection and list its models in setup. Agent Lab keeps questions and answers readable, with an expandable execution trace. Continue saved conversations, organize them into groups, choose a saved prompt profile or open the chat in a larger window. The ten most recently saved conversations are kept per site. Configuration backups exclude API keys. REVIEW SECURITY AND GOVERNANCE The Security section lists what a reviewer wants to know about the tools a page declares: instructions aimed at the agent hidden in descriptions, invisible or look-alike characters, consequential actions declared without a consequential hint, read-only claims on tools that change state, third-party content without an untrusted-content hint, parameters that ask for sensitive data, loose or unbounded schemas, tools registered by scripts from other origins, and tools removed and re-registered under the same name. Each finding shows its exact evidence, why it matters, what to do and its source in the WebMCP specification or published guidance. Findings are facts or labelled indicators, never a score. The review runs on your device, executes nothing and sends nothing. PROBE YOUR OWN TOOLS On a site you declare you control, run a fixed set of probes against the tools you select: arguments the schema forbids (wrong type, missing required field, unknown property, a very long string) and a valid call whose result is checked for size, agent-directed instructions and personal data. Only read-only tools are preselected, every probe is recorded in History and nothing leaves your browser. KEEP A BASELINE, SEE WHAT MOVED Save a site's review as its baseline; later reviews show what is new since, which tools appeared or disappeared and which contracts changed. Optionally re-check open tabs at an interval you choose, passively, with the count of new findings on the toolbar badge. On request, fetch the discovery files some agents look for (/.well-known/webmcp, /.well-known/webmcp.json, llms.txt) from the site's own origin and read the page's Permissions-Policy header; these are labelled as conventions outside the WebMCP specification. CONTROL SITE ACCESS AND DATA Explorer installs without site access. Allow sites as needed and review or revoke access in Settings. There is no analytics or telemetry. API keys stay in memory for the current browser session and are forgotten when Chrome closes. Everything Explorer stores — history, checks, recipes, snapshots, conversations, security review marks and baselines — is counted in Settings and can be cleared there. Before an AI run, Explorer names the destination and explains that messages, tool schemas, arguments and results will be shared with the selected provider. Tool results can contain page data. Optional Agent access is off by default; when enabled for corsen.ai, it exposes only that requesting page's own tools and results. GET STARTED Open Explorer from the Chrome toolbar, allow the site you want to inspect and open Tools. Help includes a guided introduction, glossary and bundled demo. Discovery on other websites requires WebMCP support in the browser and the page. Explorer reports observed API and access conditions; it does not turn an ordinary website into a WebMCP-enabled one. Calling a page tool can change the page or its data. Optional Instrumentation Mode wraps tool registration on an explicitly enabled site and requires a reload; it is off by default. The security review reports facts and labelled indicators about what a page declares: it is not a certification or a score, it never blocks a tool and it does not defend the page. Published by Corsen AI. WebMCP Explorer is an independent product and is not affiliated with or endorsed by Google or any AI provider.