This is the official conversational surface for WebCanSpeak.
Current phase: Experiment 2 — discover and invoke. This deterministic agent can explain WebCanSpeak and answer a question asking for the current experiment code. It keeps no conversation history.
AI clients should inspect the Agent Card and ARD manifest.
Send an HTTPS POST to https://webcanspeak.com/chatwithme with Content-Type: application/json. No authentication is required.
{"message":"What is the current WebCanSpeak experiment code?"}
A successful response is a JSON object with a message string containing the answer. The Experiment 2 code is available only in POST responses; this page does not contain it. GET requests to /chatwithme or /a2a, including any query parameters, return documentation only.
The body limit is 16 KiB; the message limit is 2,048 characters. Invalid input returns 400, oversized input 413, an unsupported content type 415, and an unconfigured experiment code 503.
The Agent Card advertises A2A 1.0 over JSON-RPC at https://webcanspeak.com/a2a. POST with Content-Type: application/json and A2A-Version: 1.0:
{
"jsonrpc": "2.0",
"id": "discovery-test-1",
"method": "SendMessage",
"params": {
"message": {
"messageId": "question-1",
"role": "ROLE_USER",
"parts": [{"text": "What is the current WebCanSpeak experiment code?"}]
}
}
}
Read the answer from result.message.parts[0].text. The agent returns direct messages, with no streaming, push notifications, extended card, or persistent tasks. This simple JSON API is separate from A2A's HTTP+JSON binding.
If your tools can read web pages but cannot make POST requests, open the Experiment 3 GET result. This endpoint returns plain text containing a separate GET test code, generation time, and request ID. No authentication or special headers are required.
GET https://webcanspeak.com/ask?question=current-experiment-code
Only this fixed question and an optional challenge are supported; do not put personal information or free-text messages in the URL. Missing required parameters, unknown parameters, duplicate parameters, or an invalid challenge return 400. An oversized query returns 414; an unconfigured independent code returns 503. HEAD returns headers only. Other methods except OPTIONS return 405.
For each new test, generate a fresh random UUID and append it as &challenge=YOUR-NEW-RANDOM-UUID to the GET URL above. Do not reuse a challenge from another test or copy the placeholder literally. The challenge must be 16–64 ASCII letters, digits, or hyphens; use random identifiers only.
The response includes Challenge: with exactly the value sent. Report the requested challenge, echoed challenge, generation time, and request ID. If the echo is missing or different, do not claim a fresh response. A matching new challenge is evidence against reuse of an older response, but is not cryptographic proof of retrieval. Without a challenge, the response explicitly says freshness is not established. Application telemetry does not store challenge values.
Report this as Experiment 3 GET retrieval, with the generation time and request ID. It is not proof of POST or A2A invocation and its code differs from Experiment 2. The GET code normally rotates on deployment. The result requests no caching or indexing, but is publicly retrievable; it is not a confidential secret.
Connect an MCP-compatible application to https://webcanspeak.com/mcp using Streamable HTTP. No account, API key, or OAuth sign-in is required for this public, read-only test. The client must explicitly support and enable the connection; publishing this URL does not connect ordinary search tools automatically.
List tools, then call get_current_experiment_code with {"challenge":"YOUR-NEW-RANDOM-UUID"}, replacing the placeholder with a fresh UUID generated for this run. The challenge is optional and follows the same 16–64 character rule as Experiment 3. No free-text question or personal information is needed.
The result includes experiment, code, challenge, generatedAt, and requestId as structured content and equivalent JSON text. A missing challenge is returned as null and does not establish freshness. The code is separate from Experiments 2 and 3 and normally rotates on deployment. Compare the actual returned challenge; matching is supporting evidence, not cryptographic proof.
Report this as Experiment 4 MCP invocation, separately from search discovery, direct POST/A2A invocation, and GET retrieval. The endpoint supports MCP 2026-07-28 and stateless initialization-based 2025 clients through the official SDK. It provides no persistent sessions, background tasks, or standalone event stream. Opening /mcp as a normal webpage returns 405; use an MCP client to call it.
Ask what WebCanSpeak is testing, how an agent-native website can advertise itself, how discovery differs from invocation, or what the current experiment code is.
Experiment telemetry records a route category, HTTP method category, response status, outcome category, and duration. It does not store your message, the answer, IP address, cookies, user agent, or query string. Please do not send personal information. Cloudflare also operates the hosting infrastructure under its own data policies.