mcp-home: plan

What this server does, how sign-in works and what data it keeps.

Endpoint

https://mcp-home-k7q2.pages.dev/mcp: MCP over Streamable HTTP, POST only. Protocol versions: 2026-07-28, 2025-11-25, 2025-06-18, 2025-03-26. 2026-07-28 is served statelessly (version and capabilities in each request's _meta, server/discover supported). Older clients may use the initialize handshake; no sessions are created.

Tools

Tool	What it does	Input	Output (structured content)
`memo_write`	Saves a text note on this server and returns its id.	`text`	`id, created_at`
`memo_search`	Finds notes whose text contains every word of the query (case-insensitive), newest first. An empty query returns the newest notes.	`query, limit`	`notes`
`upload`	Stores a file on this server and returns a public link that anyone can open. Give the file as `content` (text, or base64 with encoding "base64") or as a `url` to copy from. Max 10 MiB.	`filename, content, encoding, mime_type, url`	`url, id, filename, size, mime_type`
`transcribe`	Fetches audio from a URL and turns it into timestamped text with Whisper (Workers AI). start and end are seconds from the beginning. Max 25 MiB.	`url, language`	`segments, text, language, duration`

Every tool declares an outputSchema. Results come back as structuredContent that matches it, plus the same JSON as text. Bad input comes back as a tool result with isError: true.

memo_write { text } → { id }. Up to 20,000 characters.
memo_search { query } → { notes: [{ id, text, created_at, source }] }. Every word of the query must appear (case-insensitive); newest first; empty query = newest notes.
upload { filename, content, encoding? , mime_type? } or { url } → { url }. Up to 10 MiB. The link is served by this server under /files/ and anyone with it can open it.
transcribe { url } → { segments: [{ start, end, text }] }, times in seconds. Up to 25 MiB of audio; uses Whisper (@cf/openai/whisper-large-v3-turbo) on Cloudflare Workers AI.

Sign-in

OAuth 2.1 with PKCE. This server is both the authorization server and the protected resource.
Discovery: protected resource metadata (RFC 9728) at /.well-known/oauth-protected-resource/mcp; authorization server metadata (RFC 8414) at /.well-known/oauth-authorization-server. A request without a token gets 401 with a WWW-Authenticate header that names the metadata.
Clients: Client ID Metadata Documents (preferred) or Dynamic Client Registration at /oauth/register.
Scope: mcp (all four tools). Tokens are bound to the resource https://mcp-home-k7q2.pages.dev/mcp (RFC 8707) and last one hour; refresh tokens are issued.
Who gets in: only clients the owner lets in. The consent page shows which app is asking and where the access goes; it opens only with the owner password. Wrong passwords are capped at 10 per hour. Deny sends the app away with access_denied.

Data it keeps

What	Where	How long
Notes: text, time, name of the app that wrote them	Cloudflare D1	Until the owner deletes them
Uploaded files and their name, type, size, time	Cloudflare KV	Until the owner deletes them. Public to anyone with the link.
OAuth clients, grants and tokens (tokens and secrets only as hashes)	Cloudflare KV	Until they expire or are revoked
Audio sent to `transcribe`	Not stored. Fetched, sent to Workers AI (zero data retention model), then dropped. The transcript is returned, not kept.	—
Request logs	Cloudflare platform logs	Cloudflare's default retention

No analytics, no cookies except the short-lived consent cookies on the sign-in page.