MCP

notefeed has an MCP endpoint, so AI assistants such as Claude can post notes to a feed and read them back. It lives at /mcp, for example https://notes.example.com/mcp, and serves any number of feeds: the feed name is a parameter of each tool.

Tool	What it does
post_note	Posts a markdown note to a feed. Takes feed and markdown, optionally a title (see Titles), tags (a list of tags) and read_id (the read id of the feed this post creates); on notefeed.me, a password on the post that creates the feed needs the license key as license_key. Optionally attachments: a list of { name, type, data, alt } with pictures (data is standard base64, not URL-safe), posted with the note in one atomic call; write  in markdown where one goes, and one the text never refers to is added at the end. The answer lists them as attachments (id, file, url). A refusal posts nothing, and the error names the attachment. The MCP request body is at most 16 MiB: one 10 MiB picture fits, several large ones in one call do not. See Posting a note with its pictures.
list_notes	Lists a feed's notes, newest first, without their text. Takes feed, optionally limit (1–100, default 20), before and tag (only notes carrying it). Each note lists its tags.
get_note	Reads one note with its content. Takes feed and id.
edit_note	Replaces a note's markdown or sets its title (an empty title removes it), or both; its id stays. Takes feed, id and markdown and/or title, and returns the note.
delete_note	Permanently deletes a note. Takes feed and id, and returns { "deleted": true }. It is marked as destructive, so clients can ask you before running it.
get_feed	Takes feed, and returns its title, description, title image URL, whether it is protected, and its read link.
update_feed	Replaces a feed's title, description and title image, and can change its read id. Takes feed, title, description and optionally image and read_id (empty for a random one; see Choosing a feed's read id), and returns the feed.
delete_feed	Permanently deletes a feed with all its notes, settings, password and read link. Takes feed, and returns { "deleted": true }. It is marked as destructive.
post_file	Posts a file as a note, as base64. Takes feed, type (a media type from Types: text/markdown, image/png, image/jpeg, image/gif or image/webp) and data, optionally title, alt and name, and returns id, file and url (the file's public URL; write  in a markdown note to show a picture). The data must be what type says. The same rules as the API: size limit, rate limit, public URL. update_feed takes the returned file as image to make it the feed's title image.

Posting follows the same rules as the HTTP API: size limit, rate limit, caps, and the first post creates the feed. Editing and deleting follow Editing and deleting notes: the same access as posting, the same rate limit, and a note that does not exist is the error "no such note". Anyone who can post to a feed can use these tools on it, and a deleted note cannot be brought back, so think before giving an assistant a feed it should only add to. Results use the API's field names.

Protocol version

The endpoint speaks MCP 2026-07-28 and nothing older. A client that only knows older versions can't connect: it gets a 400 naming the supported version. Update the client.

Claude Code

On an open instance:

claude mcp add notefeed --transport http https://notes.example.com/mcp

On an instance with a password, send it as a bearer token, as the API does:

claude mcp add notefeed --transport http https://notes.example.com/mcp --header "Authorization: Bearer <password>"

Claude.ai and Claude Desktop

Add a custom connector and give it the URL https://notes.example.com/mcp. On an open instance that's all. On an instance with a password, the connector opens notefeed's login page: check that it names the client and the host it returns to, enter the instance password and press Allow. See OAuth for how that works and how long it lasts. If the instance has sign-in switched on, that login page also has a Sign in with button for each provider, and no extra setup is needed at the providers: the one redirect URI of the sign-in setup covers MCP clients too.

OAuth needs the instance's public https:// address: set PUBLIC_URL=https://notes.example.com, or NOTEFEED_TRUST_PROXY=1 behind a reverse proxy that sets X-Forwarded-Proto and X-Forwarded-Host.

A client may register only https:// redirect URIs, or plain http:// on localhost or 127.0.0.1, and must return to exactly the one it registered, port included. Clients whose callback has its own scheme, such as cursor://, can't log in through OAuth: send the password as a Bearer header instead, as for Claude Code.

Feeds with a password

A feed can have its own password (see Feed passwords). All nine tools take an optional password; an empty one is the same as none. post_note with a password to a feed that doesn't exist yet creates it protected; to an existing feed that has none, it answers with the error "feed already exists and has no password". On a protected feed, every tool needs the right password, or answers with the error "missing or wrong password". Wrong passwords count toward the failed-attempt limit; a call without a password does not. The password is not remembered between calls, and it is part of the conversation, so tell the assistant only passwords you would give that service. It is in addition to the instance login, which still decides who may use /mcp at all.

Telling the assistant which feed

The assistant needs a feed name. Put it in a project or custom instruction, for example "Post notes to the notefeed feed homelab-7f3k2q9x4m8wz."

The feed name works like a password

Anyone who knows it can read the feed, post to it, and edit and delete its notes, change its settings, delete the feed, and the assistant sees it. The tool descriptions ask the model not to repeat it, but don't give it a name you wouldn't share with that service. Use a hard-to-guess name.

Notes the assistant reads are text from your feed: don't point it at a feed that strangers can post to.