Install

POST /mcp is MCP Streamable HTTP. GET /mcp is 405. There is no legacy SSE endpoint. Auth on every client below is none. The anonymous free tier is 60 tools/call requests per 60 seconds per client.

Tools, in list order: score, heard_as, recognize, create_upload, score_attempt. Every tool sends title, inputSchema, outputSchema, and annotations.readOnlyHint, openWorldHint, and destructiveHint.

Chat, Claude, and Gemini pass JSON. They do not attach microphone audio to a tool call. Send ipa, audio_base64 (a WAV, at most 256 KiB decoded), or audio_url. ipa wins when it is also sent. A larger clip calls create_upload, PUTs the wav to upload_url with Content-Type: audio/wav, and passes audio_url. This build stores that wav in process memory for 15 minutes. It does not sign Cloudflare R2.

score and score_attempt point at the record widget ui://ipa/record.html (/apps/record). The widget MIME type is text/html;profile=mcp-app. Codex Desktop and ChatGPT on iOS deny getUserMedia even when ui.permissions.microphone is declared. ChatGPT on desktop Chrome can grant it.

ChatGPT

Creating custom MCP apps in ChatGPT: Plugins, then the plus, then Add custom MCP server. URL https://ipa.raymati.com/mcp. Auth none. Create as a plugin.

App review is the OpenAI plugin submission, not a self-serve listing. Review checks the HTTP Content-Security-Policy and the widget _meta.ui.csp (connectDomains and resourceDomains are https://ipa.raymati.com).

Claude

In the Claude app: Customize, Connectors, Add custom connector, URL https://ipa.raymati.com/mcp.

Directory submission is separate: <https://claude.ai/directory/manage>. Steps are in Submitting.

Claude Code

claude mcp add --transport http ipa https://ipa.raymati.com/mcp

A JSON entry has to include "type": "http". A url without type is treated as stdio and skipped. streamable-http is an alias for http. Source: <https://code.claude.com/docs/en/mcp>.

Cursor

~/.cursor/mcp.json or .cursor/mcp.json. Source: <https://cursor.com/docs/mcp>.

{"mcpServers":{"ipa":{"url":"https://ipa.raymati.com/mcp"}}}

VS Code

.vscode/mcp.json on VS Code 1.101 or newer.

{"servers":{"ipa":{"type":"http","url":"https://ipa.raymati.com/mcp"}}}

Gemini CLI

gemini mcp add --transport http ipa https://ipa.raymati.com/mcp

settings.json uses httpUrl, not url. Source: <https://geminicli.com/docs/tools/mcp-server/>.

Content-Security-Policy

HTML responses, including this page and /apps/record, send:

default-src 'self'; script-src 'self' 'unsafe-inline'; style-src 'self'; img-src 'self'; connect-src 'self'; base-uri 'self'; form-action 'self'; frame-ancestors 'self' https://chatgpt.com https://chat.openai.com https://claude.ai

'unsafe-inline' is there so the landing application/ld+json block and the inline record widget can run. Widget review uses _meta.ui.csp as well as this header.

robots.txt

ALLOW_TRAINING_CRAWLERS in scripts/build_api_docs.py is one flag. It defaults to True.

When it is true, GPTBot, ClaudeBot, Google-Extended, and CCBot get Allow: /. When it is false, those four get Disallow: /.

These stay Allow: / either way: ChatGPT-User, OAI-SearchBot, Claude-User, Claude-SearchBot, PerplexityBot, Perplexity-User, and User-agent: *. A named group wins over * for that bot.

Google-Extended and Applebot-Extended are control tokens, not HTTP user-agents. Bytespider is listed and is widely reported to ignore robots.txt.