Connect Claude to your vault

Claude can read your Research Workbench vault while you talk to it: search your papers and notes by what they mean, read a document page by page, find what is in a project, and write a citation in your style. It changes nothing by itself: it can propose a note or a reference, which waits for you to allow it - in the cloud only if you let it propose changes when you connected it.

It works through the Model Context Protocol, the standard way for an assistant to reach your own tools. Your vault stays where it is: Claude asks for a passage or a page, and the workbench decides what it may have.

Built For the workbench in the cloud, with a subscription, and for the workbench on your computer, with Claude Desktop or Claude Code. Claude first; other assistants are added one at a time, as each is tested.

Connecting Claude

Once. The same steps in Claude on the web, on your desktop or on your phone.

In Claude, open Settings, then Connectors, and choose Add custom connector.
Name it Research Workbench and give it this address: https://app.researchworkbench.space/mcp
Open Advanced settings and enter the OAuth client ID claude. Leave the client secret empty.
Choose Add, then Connect. Claude sends you here, to this site.
Sign in if you are not, and choose Allow. You are sent back to Claude, and the vault's tools appear in the chat.

The workbench in the cloud shows the same steps under Settings, Assistants, with its address to copy. Use an account that has opened the cloud workbench at least once: that is where your vault is.

On your computer

Claude Desktop or Claude Code, and the workbench open beside it. No account needed.

In the workbench on your computer, open Settings, Assistants. It gives, with this computer's own path, the entry to add to Claude Desktop's configuration (Settings, Developer, Edit Config, then restart Claude Desktop) and the one command that adds it to Claude Code.

Claude starts a small helper that passes each request to the open workbench and opens nothing itself. If the workbench is not open, Claude is told so. What it reads still goes to Anthropic, so the rules below apply just the same.

What Claude can do with it

Four tools, all of them reading. Ask in your own words; Claude chooses the tool.

Search the vault: Passages that answer a question, by their words and their meaning, each with its document, section, page and position.
Read a document: A page, or a stretch of up to 20,000 characters, with where its pages fall, so a quotation can be cited by page.
List the vault: Your projects, and a project's folders and documents.
Cite a reference: A reference from a project's library as your papers cite it, in APA 7, Harvard, IEEE or Chicago author–date.

What it is never given

The same rules as for any hosted AI model, decided by the workbench, not by Claude.

Documents labelled personal or special category data, unless their folder has a recorded exception for hosted models. Claude is told how many were left out, so you know why an answer is thin.
Anything at all while privacy mode is on, or, in the cloud, after a subscription has ended.
A way to change your vault by itself. Claude can propose a note on a document or a reference by its DOI; each waits under Approvals, showing exactly what would be added, until you allow or refuse it. In the cloud it can propose only if you ticked Also let it propose changes when you connected it.

Everything Claude is given is recorded, as every AI call is: in the workbench, Activity, Evidence lists each request, marked assistant, with exactly what was given, and Usage counts them. What your documents say is passed to Claude as material to read, never as instructions.

Stopping it

On your computer, remove research-workbench from Claude's configuration, or turn privacy mode on.

In the cloud: on this site, under your account, Assistants, every assistant you allowed is listed with when it last connected. Revoke stops it: it cannot connect again, and a request already under way has no way in after 15 minutes at most. To use it again, connect it again.

In the cloud, each person has a limit of requests a minute and a day, so a runaway conversation cannot read the vault without end; Claude is told when it may ask again.