by James M Dewar
This project is unfunded — depending on usage, rate limits may apply.

Connecting tools

Give the model sources it can search, so that it works from records retrieved for this task rather than from memory. This guide adds a literature source to ChatGPT or Claude on the web and checks that it is being used.

Steps

In the workshop, these steps are the exercise. You can also follow them on your own.

  1. Pick a literature source. Consensus works without an account, and PubMed (in Claude only) needs no sign-in; Wiley Scholar Gateway, Scite and Elicit need accounts of their own.
  2. Add it in your tool by following Step 2 · Add it in ChatGPT (web) or Step 2 · Add it in Claude (web). For any other tool, use the button above them.
  3. Start a new conversation, turn the source on, and run the prompt from Step 3 · Test prompt.
  4. Use Step 4 · Is it working? to confirm it worked. The conversation should show the call to the source, and the identifiers it returned should resolve to the papers named.

These steps use the labels in each maker’s documentation as read on 1 October 2026. Both makers rename menus often, so a label may differ in your account.

Step 2 · Add it in ChatGPT (web)

ChatGPT calls these plugins and apps (formerly connectors). Installing from the directory is the simpler route. Adding a server by its address needs Developer mode, which OpenAI lists for Plus, Pro, Business, Enterprise and Education accounts, on the web only; in a work or education workspace an administrator may have to allow it first.

From the plugin directory

  1. Select Plugins in the sidebar, or open chatgpt.com/plugins.
  2. Search for the source and open it. Scite and Consensus are listed; Elicit and Wiley say theirs are too.
  3. Select Install plugin (one OpenAI page describes it as a plus button), then Connect if asked, and sign in to your account with the source.

By address, for a server that is not listed

  1. Open Settings → Security and login and turn on Developer mode. One OpenAI page gives Settings → Apps → Advanced settings instead; look there if the first has no such switch.
  2. Go to Plugins and select the plus button.
  3. Enter a name and a description. Under Connection, paste the server’s address, including /mcp. The addresses are under For other tools.
  4. Choose the authentication if asked, complete the sign-in, and create the connection.
  5. Review the tools it discovered. If it found none, it is not connected.
  6. If it is not then offered in a conversation, open Plugins, look under Personal, and install it from there.

Then, in a conversation

  1. Start a new conversation. Select + in the message box and choose the plugin, or type @ and its name. For a server added by address, choose Developer mode from the + menu and select it there.
  2. Name the source in your request. ChatGPT otherwise decides for itself whether to call an installed plugin.
  3. If an approval card appears, read what it proposes to send, then allow it.

OpenAI's documentation

Step 2 · Add it in Claude (web)

Claude calls these connectors; one you add by address is a custom connector. Anthropic lists them for every plan, Free included, with a Free account limited to one custom connector. On a Team or Enterprise plan an Owner adds it for the organization first, and you then select Connect.

From the connectors directory

  1. Open Customize → Connectors (claude.ai/customize/connectors).
  2. Select Discover, search for the source, and open its page. Consensus, PubMed, Wiley Scholar Gateway, Scite and Elicit are all listed.
  3. Select Connect to Claude (one Anthropic page says Connect or Install), sign in on the source’s page, and approve the access it asks for.

By address, for a server that is not listed

  1. Open Customize → Connectors.
  2. Select + Add, then Add custom connector. One Anthropic page gives Add → Custom → Web for the same thing.
  3. Enter a name and paste the server’s address (the addresses are under For other tools). Select Continue.
  4. Keep the authentication settings Claude detected unless the source says otherwise, and select Add.
  5. Sign in to the source and approve the access it asks for.

Then, in a conversation

  1. In a new chat, select + in the message box, then Connectors, and check that the source’s switch is on.
  2. Ask for something that needs the source.
  3. If Claude asks to use one of its tools, select Allow once.

Anthropic's documentation

Step 3 · Test prompt

A request that tells the model to call the source, to say what it did, and to stop rather than answer from memory.

When to use it: In a new conversation, with the source turned on.

Fill in

Then open two of the identifiers or links yourself. The visible call shows the source was used; an identifier that resolves to the paper named shows the record is real. You need both.

Copy this

Use the [name of the source, e.g. Scite] connector to find three peer-reviewed papers on: [your topic]

For each paper, give the title, the year, and the DOI, PMID or link exactly as the connector returned them. Then state which tool you called, the query you sent, and whether you received abstracts, citation contexts, or full text.

If the connector is not available in this conversation, or returns nothing, say so and stop. Do not answer from memory.

Step 4 · Is it working?

A fluent answer is not evidence that the source was used. Look for these.

It is connected

  • Claude: the source is listed under Your connectors as Connected, and its page lists its tools under Tool permissions.
  • ChatGPT: the plugin is listed under Settings → Plugins. A server added by address showed you the tools it discovered when you created it.
  • Connected is not the same as in use. In Claude the source’s switch has to be on for the conversation. In ChatGPT an installed plugin is called when ChatGPT judges it relevant or when you select it (+ or @), and a server added by address has to be selected under Developer mode in the + menu.

It was used for this answer

  • Claude may ask before using a tool (Allow once, Always allow). No request is not a sign of failure: check for a result that came from the source.
  • ChatGPT may show an approval card first. In Developer mode each tool call can be expanded to show the input that was sent and the output that came back.
  • The identifiers or links (DOI, PMID) in the answer match what the call returned, and resolve to the papers named.
  • No sign of a call means the answer came from the model’s memory, however fluent it is.

Inspect the result

  • Does the conversation show a call to the source and the query that was sent?
  • Do the returned DOIs or PMIDs resolve to the papers named?
  • Did the model say what it received: abstracts, citation contexts, or full text?
  • Did it say so when the source returned nothing, rather than answering from memory?

Record

Which source you connected, in which tool, and what the test returned.

When to use this

The answer depends on the literature and you need to be able to recover the source behind each claim.

Continue to Summarize and restart