Skip to content
Octane AIOctane AI

Connect an assistant

Octane AI runs an MCP server at https://mcp.octaneai.com/mcp. Paste that address into any assistant that speaks MCP, sign in to Octane AI, choose the store and the permissions, and the assistant can read your quizzes, ask for analytics, look people up, start exports and build quiz drafts with the editor’s own assistant.

No API key is involved. The server takes only the sign-in you complete in the browser; a key from Developer > API keys does not work on it. The REST API and its keys are a separate path, described under Authentication.

The sign-in

The first time an assistant connects, it sends you to the Octane AI sign-in. The flow is the same in every client:

  1. Sign in to your Octane AI account. The page asks “Connect this assistant?” and waits for your Connect; Cancel returns you to the dashboard and grants nothing.
  2. Pick the store. An account that administers several stores also offers All my stores, which covers every store you administer, including stores you take on later.
  3. Pick the permissions, one choice per area:
    • Quizzes: read, write, publish, or no access. Write lets the assistant edit drafts and archive or unarchive a quiz; publish lets it publish a version.
    • Analytics: read or no access.
    • Shopper profiles: read or no access. This is the one permission that exposes personal data (names, emails, phone numbers and answers).
  4. Confirm. The assistant receives a token carrying exactly those choices and nothing more.

The connection belongs to your account, not to one store. It appears in the dashboard under Developer > Limits as a connected app on every store you administer, with a Disconnect button (the Limits tab). Disconnecting there revokes the grant: the assistant gets no new token, and a session already open stops working when its current token expires.

Claude on the web, mobile and desktop

Claude adds a remote MCP server as a custom connector. Anthropic’s help center covers every plan: Free (one custom connector), Pro, Max, Team and Enterprise (Get started with custom connectors, updated August 2026).

  1. On claude.ai or in Claude Desktop, open Customize > Connectors.
  2. Click +, then Add custom connector.
  3. Name it “Octane AI”, paste https://mcp.octaneai.com/mcp as the URL, and click Add. Leave the advanced OAuth fields empty; the server registers the client itself.
  4. Click Connect and complete the sign-in.
  5. In a chat, open the + menu, then Connectors, and switch Octane AI on for that conversation.

On a Team or Enterprise plan an Owner adds the connector first under Organization settings > Connectors > Add > Custom > Web; each member then connects it from Customize > Connectors.

Claude Desktop uses the same connector list as the web: the connection is brokered through your Claude account and reaches Octane AI from Anthropic’s servers, not from your computer, so nothing is installed locally and no local config file is involved. The mobile apps use the connectors you added on the web; a custom connector cannot be added from the phone (Use connectors).

To disconnect, open Customize > Connectors, find Octane AI and choose Remove from its menu. To change permissions, remove it and add it again, then pick the new permissions at sign-in.

Claude Code

Claude Code adds remote HTTP servers from the command line and signs in from the /mcp panel (Connect Claude Code to tools via MCP).

Terminal window
claude mcp add --transport http octane https://mcp.octaneai.com/mcp

Add --scope user to make the server available in every project instead of the current one. Then, inside a Claude Code session:

/mcp

Pick octane, choose Authenticate, and complete the sign-in in the browser. The server shows connected in the /mcp menu once it is done. Claude Code refreshes the token on its own; when a refresh fails, /mcp offers Re-authenticate.

To disconnect, run claude mcp remove octane (add --scope user if you added it that way), then revoke the app in the dashboard.

ChatGPT

ChatGPT connects custom MCP servers through developer mode, on the web only. OpenAI’s platform guide lists Plus, Pro, Business, Enterprise and Education accounts (ChatGPT developer mode); its help center adds that full support including write actions is in beta for Business, Enterprise and Edu workspaces, and that Pro accounts get read and fetch actions in developer mode (Developer mode and MCP apps in ChatGPT, updated August 2026). Expect reads and analytics to work everywhere; count on drafts, publishing and archiving only on a workspace plan.

On a personal plan:

  1. Open Settings > Security and login and turn on Developer mode.
  2. Open Settings > Plugins (or go to chatgpt.com/plugins) and click + to create a developer-mode app.
  3. Name it “Octane AI”, enter https://mcp.octaneai.com/mcp as the server URL, choose OAuth as the authentication, and complete the sign-in when ChatGPT opens it.
  4. In a new chat, choose Developer mode from the + menu and select the Octane AI app for the conversation.

On a Business, Enterprise or Edu workspace an admin enables developer mode in the workspace settings, creates the app under Workspace settings > Apps > Create (enter the URL, choose OAuth, click Scan Tools, sign in, then Create), tests it as a draft and publishes it to the workspace.

Developer mode does not require the search and fetch tools that deep research and company knowledge look for, so every Octane AI tool is available in an ordinary chat. The Octane AI server has no search or fetch tool, so it does not take part in deep research or company knowledge. ChatGPT treats every tool the server marks read-only as a read; the others (drafts, publish, archive, exports) ask for your confirmation before they run.

To disconnect, open the app under Settings > Plugins and remove it; on a workspace plan an admin removes it under Workspace settings > Apps. Then revoke the app in the dashboard.

OpenAI Codex

Codex stores MCP servers in ~/.codex/config.toml, shared by the CLI, the IDE extension and the ChatGPT desktop app (Model Context Protocol in Codex).

From the terminal:

Terminal window
codex mcp add octane --url https://mcp.octaneai.com/mcp
codex mcp login octane

codex mcp login opens the browser for the sign-in and stores the token in your keyring; /mcp inside the Codex TUI lists the connected servers. When a token stops working, run codex mcp logout octane and codex mcp login octane again.

In the IDE extension, open the gear menu, choose MCP servers, click Add server, name it, choose Streamable HTTP, enter https://mcp.octaneai.com/mcp, save, and click Restart extension; the list marks the server as needing OAuth and offers Authenticate. The ChatGPT desktop app has the same list under Settings > MCP servers with Add server and Restart. If the app does not offer a sign-in button for the server, codex mcp login octane from the terminal signs it in for every Codex surface.

To disconnect, run codex mcp logout octane and codex mcp remove octane, then revoke the app in the dashboard.

Cursor

Cursor adds MCP servers from the Customize page in the sidebar or from a JSON file; remote servers with OAuth are supported (Model Context Protocol in Cursor).

Create .cursor/mcp.json in a project, or ~/.cursor/mcp.json for every project:

{
  "mcpServers": {
    "octane": {
      "url": "https://mcp.octaneai.com/mcp"
    }
  }
}

No auth block is needed; the server accepts dynamic client registration. Cursor picks the file up, and the first connection opens the browser for the sign-in. Cursor asks before running an MCP tool by default; expand the call to see its arguments.

To disconnect, delete the entry from mcp.json (or remove the server on the Customize page), then revoke the app in the dashboard. The toggle on the Customize page disables the server without removing it.

The first conversation

Two prompts prove the connection:

Which stores can you see?

The assistant lists the stores the connection covers. On an all-stores connection, name the store in every request from then on (“for the Acme store”); the assistant passes it to every tool.

How did my quizzes do last week?

The assistant asks for the store-wide overview and the per-quiz summary and reports starts, completions, revenue and the like. From there, ask what you would ask a colleague: “Which answers to the skin-type question lead to the most orders?”, “Draft a version of the holiday quiz with a fourth product page”, “Export last month’s responses”.

What the assistant can and cannot do

The server exposes fourteen tools: list stores, read a store, list quizzes, read a quiz (its pages, questions, options, rules and dimensions), list its versions, publish a version, archive a quiz, ask analytics (the same views as Analytics), find people, read a profile, start an export and check on it, and start a quiz build and check on it. The reference describes the data each one answers with; a tool answers the same shape as its REST route.

  • Reads and analytics run as soon as you ask, within the permissions you granted.
  • Writes go to drafts. A quiz build runs the editor’s own assistant on the quiz (or creates a new draft quiz first) and saves its work as a draft; you find the conversation in the editor’s AI sidebar. Nothing a build does reaches your storefront.
  • Publish and archive ask for confirmation. Publishing needs publish permission, archiving needs write permission, and both need an explicit confirm from you in the conversation; the assistant asks before it sends one.
  • Shopper data stays out unless you granted profiles access. Without it the people and profile tools are refused, and the assistant sees no names, emails or phone numbers in a conversation. An export is a file, built in the background and downloaded by you; the assistant only starts it and reports when it is ready.
  • Every call counts against the store’s rate limits like a REST request and shows in Developer > API logs.

Troubleshooting

  • The assistant says a permission is missing. The server refuses a tool the grant does not cover and names the permission. Disconnect the app in the dashboard, remove the connector in the client, add it again and grant the missing permission at sign-in.
  • The assistant asks which store. On an all-stores connection every request needs the store named; the server tells the assistant to list the stores first. Say the store’s name as the dashboard shows it, or connect again and pick a single store.
  • The connection stops working after a while. Tokens expire. Claude refreshes them and prompts you to reconnect from Customize > Connectors when it cannot; Claude Code offers Re-authenticate in /mcp; Codex needs codex mcp login octane; ChatGPT and Cursor open the sign-in again on the next request.
  • ChatGPT lists no Octane AI tools. Developer mode is off, or the app was added as a deep research source. Turn developer mode on under Settings > Security and login and select the app from Developer mode in the composer.
  • The assistant says the quiz is being edited. A build is already running on that quiz, from this connection or from the editor. Wait for it to finish, then ask again.
  • You disconnected in the dashboard but the assistant still answers. The token it holds keeps working until it expires; remove the connector in the client to stop it at once.