NoteMCP
Save thoughts, links, screenshots and voice notes with transcriptions in a couple of seconds — no folders needed— and whichever AI you already use can search…
Что умеет
- Search Notes: Search the user's notes. Words are matched against the note's text, its tags, and the title, description and URL of any link saved with it. The default mode ("auto") escalates only as fa
- Get Note Asset: Return an image already associated with a stored note as native MCP image content. Use this after search or list when visual inspection would help identify or understand a candidate so
- List Recent Notes: List the user's notes, most recently touched first — the same order as the app's stream, so "recent" means the same thing in both. Returns thread heads only unless you ask for conti
Какие данные видит
Нужен ли аккаунт
Не нужен: сервер работает без входа
Save thoughts, links, screenshots and voice notes with transcriptions in a couple of seconds — no folders needed— and whichever AI you already use can search and work from them later.
It's for people who capture more than they organise: founders keeping decisions, builders keeping bug repros and prompts, designers keeping references, and anyone who saves a photo of a chair and wants to find it eight months later by describing it.
What the agent can do • search_notes — keyword, meaning-based, or both, filtered by tag, source, date or thread • list_recent_notes — the stream, most recently touched first • get_note — one note in full, with the thread attached to it • get_note_asset — an image already on a note, resized and compressed on request • list_tags — every tag with a live count • create_note, append_to_note, replace_in_note, update_note — write a new note, add to one, change a single span, or replace whole fields • tag_note, untag_note, delete_note
What it can't touch • Note bodies never appear in list or search replies — those carry a label and a short preview, so a listing can't flood the context window. • Embedded photos are redacted to ![photo] even in a full note. Retrieving one is an explicit get_note_asset call returning a bounded, resized copy, and a whole-body rewrite of a note containing a photo is refused, so an agent can't destroy an image it can't see. • Every write accepts an optional if_updated_at. A note changed on the phone mid-edit produces a rejected write, never a silent overwrite. • Deletes are soft — the note is hidden, not erased. • A token only ever reaches the notes of the person who authorised it, and revoking it from the account page kills the connection.
Free to start. https://notemcp.com
Список инструментов сервера (12)
Технические названия из tools/list. Нужны только разработчикам.
| search_notes | Search the user's notes. Words are matched against the note's text, its tags, and the title, description and URL of any link saved with it. The default mode ("auto") escalates only as far as it needs to: every word, then any word, then meaning — and the reply's `matched` field says which one answered, so a relaxed or semantic result is never mistaken for an exact one. Pin the strategy with mode="keyword" (every word must appear; never widens), "semantic" (meaning only) or "hybrid" (both, fused). Source, date, photo and thread filters compose with the text query and tags. Filter by tag with `tags`, which is the right tool for a request like "my #bug notes". An empty result carries `hints`: what the collection actually contains and which query to try next — read them instead of guessing another wording. Bodies are omitted unless you pass full=true; complete imported source metadata is omitted with them. Compact notes come back as { id, label, user_text, tags, source, root_id, parent_id, is_thread_head, thread_count, updated_at, … }. `user_text` is only text deliberately written by the user; imported page or social-post copy is source metadata. `label` names the note even when it has no title — use it when showing a note to a human. `updated_at` is what you pass back as `if_updated_at` when you write. |
| get_note_asset | Return an image already associated with a stored note as native MCP image content. Use this after search or list when visual inspection would help identify or understand a candidate source; do not call it for every note. It never accepts a URL. The default asset is the note’s stored source preview; asset="body" can select an existing uploaded image referenced by the note. Images are bounded, resized, and compressed on demand, while search/list responses remain text-only. |
| list_recent_notes | List the user's notes, most recently touched first — the same order as the app's stream, so "recent" means the same thing in both. Returns thread heads only unless you ask for continuations. Bodies and complete imported source metadata are omitted unless you pass full=true. Compact notes come back as { id, label, user_text, tags, source, root_id, parent_id, is_thread_head, thread_count, updated_at, … }. `user_text` is only text deliberately written by the user; imported page or social-post copy is source metadata. `label` names the note even when it has no title — use it when showing a note to a human. `updated_at` is what you pass back as `if_updated_at` when you write. |
| get_note | Fetch one note in full, plus `thread`: the continuations appended to it, oldest first. `label` identifies it, `user_text` is user-authored, and `source` is the single canonical home for imported context. `source.image_available` tells you whether get_note_asset can fetch its preview. Body images appear as `![photo]`; retrieve bytes with get_note_asset. A note with has_photos=true cannot be rewritten whole, so patch it with replace_in_note or append_to_note. |
| list_tags | Every tag the user has, with how many live notes carry it. Read this before guessing a tag name. Tags are paths: a "project/type" name like "notemcp/bug" is listed alongside the broader "notemcp", and searching the broader one finds everything under it. |
| create_note | Create a note. Pass `parent_id` to continue an existing thread instead of starting a new note — threads are flat, so continuing a continuation continues the thread it belongs to. Pass a `client_id` uuid you generate to make the call idempotent: retrying with the same client_id returns the note the first attempt created rather than a duplicate. |
| replace_in_note | Change one span of a note without resending the rest of it. `find` is literal text, not a regex or a pattern. Ticking a checkbox is find "- [ ] Buy milk", replace "- [x] Buy milk". If the text appears more than once the call is rejected rather than guessing — add surrounding text or pass all=true. Prefer this over update_note for any edit smaller than the whole note. |
| append_to_note | Append markdown to the end of an existing note's body. |
| update_note | Replace whole fields on a note. Omitted fields are left unchanged; to empty a field, name it in `clear` (passing "" is not enough). For anything short of a full rewrite use replace_in_note — it cannot mangle the text it is not touching. Rewriting the body of a note that embeds a photo is refused, because MCP only ever showed you `![photo]`. |
| tag_note | Add tags to a note. Tags already on the note are kept. |
| untag_note | Remove tags from a note. Tags not on the note are ignored. |
| delete_note | Move a note to the trash (a soft delete — the app can still recover it). A note with continuations takes its whole thread with it, so that has to be asked for with cascade=true. |