MCP
Connect an AI over MCP and it gets ten tools for your knowledge layer. This page covers what each one does, the limits that apply, and how a typical session flows.
What connecting gives you
Once Contextick is connected, your AI has ten tools. You never call them yourself — you ask in plain language and your AI picks the right one. "Keep this for later" gets saved; "what did we decide about the pricing page?" brings it back.
Everything is scoped to your account. Each person gets their own private store, and an AI can only reach it while you have it connected.
A typical flow
- Save it. "Remember this migration plan." Your AI shapes it into a title, a self-contained note, and a short summary, then calls Save.
- Bring it back. Days later, in a different conversation — maybe with a different AI — you say "use the migration plan we wrote." Your AI calls Search and gets back the closest items as summaries.
- Read the full text. To quote it or change it, your AI calls Get with that item's id. Search results are summaries, so this is the only way to read the whole note.
- Save the revision. Your AI calls Save again with the entire updated note — the whole thing with the change applied, not just the changed part.
- Clean up. Only after that save succeeds does it call Delete on the old id. Deleting first would lose the original.
The ten tools
Save
Keeps something worth coming back to — a code snippet, an idea, a design or a decision, a report, research notes. Your AI shapes it first into a title, a self-contained note, and a short summary, and is told to keep the content faithful rather than trimming it to fit.
Takes: a title, the note, and a summary.
Returns: the new item's id, plus how much you have stored in total.
Search
Brings back what you saved earlier. Your AI passes a topic, a question, or a few keywords, and gets the saved items closest in meaning — not just ones that share a word with the query.
Each result comes back as its summary, along with when you saved it, who saved it, and — for anything saved since we started recording it — which AI it came from. That last part is the point of a shared layer: a note you saved from ChatGPT comes back in Claude with its origin intact.
Takes: what to recall, and optionally how many results.
Returns: id, title, summary, how many times it's been opened, and how it was saved — when, by whom, and which AI it came from.
Get
Opens one saved item in full. Your AI uses it to quote a note exactly, or to read the whole thing before revising it. Since search results are summaries, this is the only way to reach the full text.
Get counts as a write. Opening an item raises that item's open count, which is what makes Least opened and Most opened meaningful. Because the tool reports itself honestly as a write rather than a read, some AI clients ask you to confirm before write actions run, and Get is one of them. (Some clients, ChatGPT among them, confirm before saving too.) Nothing is changed except the count.
Takes: an id.
Returns: the item's full text.
Delete
Forgets saved items by id — one or several at a time. This is permanent and can't be undone, so the tool tells your AI to confirm which items with you before it deletes anything. When it's replacing an item with a revised version, it's told to save the new one first and delete the old id only after that succeeded.
Takes: one or more ids.
Returns: what was deleted, what wasn't found, and how much you have stored afterwards.
Usage
Answers "how much have I saved?" — the number of items in your knowledge layer and the total space they take up.
Takes: nothing.
Returns: item count and total size in bytes.
Least opened
Lists saved items by how few times their full text has been opened, fewest first — a way to see what's been sitting quietly. It sorts, it doesn't filter: if you've saved five items and all five get opened constantly, it still returns those five. Each item comes with its actual open count.
The count only tracks full-text opens. It goes up when Get runs — nothing else. If your AI found what it needed in a Search summary and answered from that, it never called Get, so the item stayed at zero even though it did its job. A low count doesn't mean unused; a well-written summary can keep an item useful without anyone ever opening it. Read the list alongside the titles and when each item was saved rather than treating the number as the whole story.
Takes: optionally how many to list.
Returns: id, title, open count, whether Search can find the item yet, and how each item was saved — when, by whom, and which AI it came from. To keep the list light this one carries no summaries; Search is where those live.
Most opened
The same list from the other end — the items whose full text you keep coming back to. These are the ones an AI needed in detail, to quote or to revise, so they tend to be the notes you actually lean on. The reverse doesn't hold, though: an item missing from this list may simply have been useful as a summary.
Takes: optionally how many to list.
Returns: the same fields as Least opened, sorted by open count the other way.
Changes
Lists what has been saved since a point in time — "what did I add since Monday?" — rather than what a note is about. Your AI gives a date and time on the first call, then follows the cursor it gets back to walk through the rest.
Results are always oldest first, and they cover additions only: an item you deleted later is not reported, and an item that finishes processing does not turn up a second time. To have your own system told as things happen instead of asking, webhooks push those events to an endpoint you own.
Takes: a starting time or a cursor, and optionally how many to list.
Returns: each item's id, title and summary, when it was saved and from where, whether it is searchable yet, plus a cursor to carry on from.
Get group
Lists the parts of one uploaded file. A document you upload from your dashboard is stored as numbered parts, and every item that came from one carries the same group label — so when a search turns up part 3 of a report, this is how your AI shows you the rest of it.
The file's title is given once for the whole group, and each part comes back as an id and a state, so you can see at a glance how much of a large upload is searchable yet. Listing a file does not open its parts, so it does not affect how often they count as opened.
Takes: the group label, which the other tools return on any item that came from an upload.
Returns: the file's title, how many parts it has and how many are searchable, still processing or failed, plus each part's id and state.
Delete group
Forgets a whole uploaded file at once, rather than the one part you happened to be looking at. Every part carrying that group label goes together, so a 77-part PDF does not have to be deleted piece by piece.
It is permanent and cannot be undone, and it is all of the file or none of it — an interrupted delete leaves the file whole rather than half gone. Notes you saved directly, and parts of other uploads, carry different labels and are left alone.
Takes: the group label.
Returns: which parts were removed, and how much you have stored afterwards.
Limits
| What | Limit | Past the limit |
|---|---|---|
| Note content, per save | 16,000 characters | Rejected — your AI splits it into coherent parts and saves each one, so nothing is dropped |
| Title | 120 characters | Trimmed to fit, with an ellipsis |
| Summary | 300 characters | Trimmed to fit, with an ellipsis |
| Search results | 1–10 (5 by default) | Brought back into range |
| Least/Most opened results | 1–20 (10 by default) | Rejected |
| Changes results | 1–100 (50 by default) | Rejected |
| Search, Least opened, Most opened, Changes from a time — per hour / per day | 60 / 300 (Free) · 60 / 600 (Pro) | Refused until the window resets; the error says how long to wait |
| Save — per hour / per day | 60 / 300 (Free) · 200 / 1,000 (Pro) | Refused until the window resets; nothing already saved is affected |
| Get, Usage, Delete, Get group, Delete group, Changes from a cursor — per hour | 600 | Refused until the hour resets |
| Storage on the Free plan | 5 MB — hundreds of pages | Saves are refused until you free up space |
Titles, summaries, and note content can't be left blank — an empty one is refused rather than filled in by guesswork. See pricing for what the Free plan includes.
Connecting
There are two ways in, and which one you use is decided by the AI you're connecting, not by you.
- Sign-in (OAuth). You approve the connection with a browser sign-in — no API key to copy. In a chat app you add Contextick as a connector in its settings; in Claude Code or Codex CLI it's a one-time sign-in command. This is how Claude, ChatGPT, Replit, Claude Code, and Codex CLI connect.
- API key. You put the Contextick endpoint and an
Authorizationheader into the tool's MCP config file. Keys are issued from your account page. This is how Cursor and Devin (formerly Windsurf) connect.
Either way it's the same account and the same private store — connect two AIs and they see the same knowledge. Step-by-step guides for each are in connection guides.
Your data
What you save goes into a database of your own — not a shared table with everyone else's notes. You can delete individual items or the entire database at any time from your account page, and disconnecting an AI cuts off its access immediately. The full details are in our privacy policy.
More
To put a file you already have straight into your knowledge layer, see document upload. To save from your own scripts, see the REST API. Common questions are answered in the FAQ. Plans and limits are on pricing. If something's unclear or broken, get in touch — a real person reads every message.