Connecting connectors

Claude

The claude.ai website and desktop application use the same procedure. It is not yet possible to add a connector from the mobile application.

1. Open connector settings

   In Claude’s left-hand panel, click Customize (“Customize”), then go to Connectors.

   

2. Click Add → Add custom connector

   The button Add is in the upper-right corner; it opens the connector-addition menu. The ready-made cards (Slack, Microsoft 365) are not needed.

   

3. Enter the name and address

   Enter the name in Latin characters; we suggest . In the field Remote MCP server URL enter the address .

   Leave the fields in the Advanced settings section blank: the OAuth Client ID and Client Secret fields are intended for technical client parameters, not your password or code. Click Add.

   

4. Click Connect

   

5. Sign in

   

   

6. Choose permissions for the tools

   Open the connector card and settings Tool permissions. There are Read-only tools three options available:

   • Always allow: allow requests without separate approval.
   • Needs approval: ask for your consent before each request.
   • Blocked: prohibit requests.

   This setting controls whether a call is permitted, not how often it occurs: Always allow does not mean that the connector will be called in every response.

   

7. Enable the connector in the required chat

   Click + next to the input field → Connectors and enable . If it is not needed in this conversation, disable it in the same place.

   Check. Open a new chat and ask: «» The response should contain case details and links to the source.

ChatGPT

Two methods are available: ask the ChatGPT agent to connect everything itself or complete the form manually. Configure it on a computer in a browser.

Method 1. Ask Work or Codex

Copy the request into a new ChatGPT chat in “Work” mode or into a Codex task. During testing, the agent began the setup and stopped at account sign-in; completion of the connection by this method has not yet been confirmed.

If sign-in is required, complete it in the ChatGPT window that opens. Enter the password or code only on the service’s sign-in page; do not send it in the chat.

Method 2. Connect manually by address

1. Open the “Plugins” section

   In a browser, go to chatgpt.com/plugins and click “Add”. If creating your own MCP is unavailable, enable developer mode in the account settings.

   

2. Select “Create MCP app”

   The “Create plugin” option opens a different process and is not needed.

   

3. Enter the address and sign-in details

   Name: . In the “Connection” section, leave “Server URL” and enter . In the “Authentication” section, select ; leave the advanced settings blank.

   

4. Confirm creation

   Read the warning about a third-party MCP server, select “I understand and want to continue” and click “Create”.

   

5. Sign in

   In the connection window, click the button to proceed to the service.

   

   

   

6. Select a connector in the chat

   In a new conversation, click “+” next to the input field and select the app or type @ followed by its name. The selected app will appear above the field.

   

   Check. Ask: “» The response should contain case details and links to the source.

Grok

The grok.com website or desktop app. It is not yet possible to add a connector from the mobile app. Connectors are also available on Grok’s free plan.

1. Open “Skills and connectors”

   In the left panel, click “Skills and Connectors” under “Automations.”

   

2. Click “New connector”

   The “Connectors” tab and the button in the upper-right corner. The cards for Gmail, Google Drive, and other services are not needed here.

   

3. Select “Custom settings”

   The first item labeled “Add your own custom connector.”

   

4. Fill in two fields

   Name in Latin characters: . Server URL: . Click “Add Connector”; Grok determines the OAuth parameters itself.

   

5. Sign in

   

   

6. Enable the connector in the conversation

   Unlike Claude, in Grok the connector is specified for the conversation: while typing your question, select it directly below the input line . The quick way: type @ and select the connector from the list. The selection remains active until the end of the conversation; you need to repeat it in a new chat.

   

   Check. With the connector selected, ask: “»

Hermes

The Nous Research agent with any model of your choice. Hermes handles the connection itself: you write one sentence containing the address, and it adds the server, opens the sign-in window, and enables the tools. The conversation takes about ten minutes, with your part taking about one minute.

Installation

If Hermes is already installed, proceed to the connection step. The installer is available on the official Hermes website; from the terminal:

iex (irm https://hermes-agent.nousresearch.com/install.ps1)   # Windows, PowerShell
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash   # macOS, Linux, WSL2

Connect with a single command in the chat

1. Enter the command in a new conversation

   Without configuring anything in advance, send a sentence containing the connector address:

   The word after “as” becomes the connector’s name in your Hermes. The name can be anything, but must use Latin characters only.

   

2. Authorize the connection

   Hermes will hermes mcp add and stop at the confirmation step Authorize the … MCP server? Click Authorize.

   

3. Sign in

   At the bottom of the sign-in page, you can see whom you are granting access to: the app Hermes Agent and the callback address on your own computer (127.0.0.1). That is how it should be.

   

   

4. Enable the connector

   Return to Hermes. It will check the connection (hermes mcp test) and ask Enable the … MCP server? Click Enable.

   

5. Check the report and start a new conversation

   At the end, the agent displays the name, address, sign-in method, “enabled” status, and number of tools found. The tools appear in new conversations: open a new one or run /reload-mcp.

   

ZCode

Unsupported environment: it works according to the MCP standard, but we do not guarantee the stability of a particular version.

1. Open the list of MCP servers

   Settings → MCP Servers → New MCP Server. Scope: User, so that the connector works in all projects.

2. Fill in the form

   Type: HTTP (not stdio or SSE). URL: . Name: .

3. Sign in

   

   

4. Test the connection

   Ask: “» The first search takes about 30 seconds.

AnythingLLM

Unsupported environment. There is no sign-in window here: access is passed in a header.

2. Open Agent Skills → MCP Servers

   Settings gear → Agent Skills; at the bottom, the MCP Servers section with a Refresh button.

   

3. Describe the server in a file

   File storage/plugins/anythingllm_mcp_servers.json. Type is required streamable, with sse the connection will not be established.

4. Click Refresh

   The Running status means that the connection has been established. Stopped: invalid key or sse type.

   

5. Ask via @agent

   The tools work only in agent mode: start the message with @agent.

   

DeepSeek Harness

Unsupported environment, for use in the terminal. There is no sign-in window: access is passed in a header.

2. Add the server to cordis.patch.yml

   File ~/.dsh/profiles/web/cordis.patch.yml. The entry must use insert, otherwise dsh will silently skip it.

3. Check the configuration

   dsh --profile web --dump-config | tail -20

4. Run dsh web

   Settings → Plugins → Plugin list; search for mcp: the mcp-client entry must be Enabled.