Can I search my docs from ChatGPT, Claude or Cursor without building RAG?

Yes. No vector database, no pipeline. Gryz search finds the right passage across your docs, notes, memories and tasks from ChatGPT, Claude or Cursor.

Short answer: Yes. Put your documents in Gryz and connect ChatGPT, Claude or Cursor over MCP. Their search tool finds the right passage across your documents, notes, saved memories and tasks, by meaning and by exact words. Every result links to its source, and when nothing matches, it says "No grounded source found" instead of guessing.

What you'd build yourself, and what Gryz does

RAG step	Building it yourself	In Gryz
Get files in	Write a loader per file type	Import `.md`, `.html`, `.txt`, `.docx` and `.odt`, up to 20 at a time, or write in Gryz
Split into passages	Pick a chunk size and test it	Done for you, along headings
Search by meaning	Run an embedding model and a vector database	Done for you
Search by exact words	Add a keyword index next to it	Done for you, in the same search
Say where it came from	Track sources through the pipeline	Every result has a title, link, heading and snippet
Say when it doesn't know	Tune a cutoff and hope	A plain signal: likely, uncertain or unlikely
Who can see what	Build permissions into retrieval	Results only come from what you can already open
Connect it to your tools	Write an integration per tool	One MCP connection works in ChatGPT, Claude, Cursor and Codex

Steps

Connect your tool. Add https://www.gryz.ai/api/mcp in Claude, Cursor, Codex or ChatGPT and sign in. Every client is in MCP setup.
Bring your documents in. On your Documents page, choose + New, then Import files. Sort them into folders, up to three levels deep.
Ask a question only your documents can answer. For example:

Search my Gryz for what we agreed with the contractor about the deposit. Quote the passage and link the source. If Gryz says no grounded source was found, tell me that instead of guessing.

Your tool calls search, reads the best results, and can call fetch to open the full text or just one section.

What gets searched

Documents. Markdown and HTML, the ones you wrote or imported.
Notes. Your private working notes.
Saved memories. Only ones that really match the question, and only if the connection could already read them with recall.
Tasks. Matched by the words in their title and description.

You can switch notes, memories and tasks on or off for search. Documents are always included.

Never searched: documents with a PIN, documents with an expiry date, documents held for review, and anything you leave out yourself.

What a result looks like

search returns up to 10 results, best first, from all four kinds together. Each one has:

a title and a link back to the source
its type: document, note, memory or task
the heading it sits under, a short snippet and a one-line summary
when it last changed, and why it matched

Document and note results also say how strong the match is: strong, partial or weak. The whole search gets one signal:

Likely: at least one of the first few results is a strong match.
Uncertain: partial matches only. Read them before relying on them.
Unlikely: only weak matches. Your tool is told to say it couldn't find the answer.

If nothing is close enough, there are no results and it says "No grounded source found". Search for a document's exact slug or link, and that document comes first.

The full contract is in the MCP tools reference.

What happens to your question

Gryz doesn't keep it. Gryz doesn't store your question or the text of the results. It records that a search happened, for cost and reliability. In a company that turns on item records, it also keeps the ids of what an app's search returned, never their content (what is private in a company).
Its words go to Cloudflare so they can be matched by meaning, the same way your passages are prepared. Cloudflare says it doesn't train on your content.
Search never makes anything public. It returns only what the person, or the agent acting for them, could already open.

Details, in plain words: How search works and where your data goes.

Settings you control

On the Knowledge page, under Settings:

Make this searchable. Turn it off and Gryz deletes your searchable passages. Your documents aren't touched.
Leave things out. Skip a single document, note or project.
Standard or Strict. Standard returns anything related, so you rarely miss an answer. Strict returns fewer, closer matches.
Try a search. Runs the real search as you, and shows why each result came back, so you see what your agents get.

Limits, stated plainly

Text only. Images and other file types aren't searched. Images, comments and footnotes in a Word file don't come across on import.
No PDF import yet. Save as .docx or .txt first.
Very long documents are searchable up to a limit, so only the first part may be included.
A match isn't proof. A passage can be on topic without holding the answer. That's why every search says how likely it is that the answer is really there.
ChatGPT needs a plan with custom MCP. As of November 2026, OpenAI's custom MCP page lists Business, Enterprise and Edu for full use, and Pro for read and fetch only. Claude allows a custom connector on every plan, Free included (Anthropic).

FAQ

Is this RAG? Yes, it's retrieval: Gryz finds the passages and your AI tool writes the answer from them. You just don't build or run any of it.

Does Gryz write the answer? No. Your tool writes it. Gryz returns the passages, their sources and how strong the match is.

Can my team search company documents? Yes, in a company workspace, once an owner or admin turns search on. A company search never returns anyone's private documents, and people only get what they could already open.

Can I search one folder? Yes, when folders are on for your account. Ask your tool to search a folder by name, like "Handbook/Policies". Notes, memories and tasks aren't in folders, so they're left out of a folder search.

Is it free? Yes, with a free Gryz account. See pricing.