Connecting Marqii MCP
Marqii MCP lets you connect your Marqii account directly to your favorite AI platform, such as Claude, ChatGPT, and more. Once connected, you can ask your AI questions about your business and take action on your Marqii data without ever logging into a separate app.
This guide walks you through how to get set up on the most common AI platforms. The first time you connect, you'll simply be prompted to sign in to your Marqii account to authorize the connection and then you're good to go!
Server facts
Transport: Streamable HTTP (no stdio, no SSE)
Auth: OAuth 2.1 authorization code + PKCE; dynamic client registration (RFC 7591) at /register
Sign-in: your normal Marqii account, via the UI at auth.marqii.com
If your brand supports remote MCP servers, you usually only need to paste the URL — the OAuth dance and client registration happen automatically.
1. Claude — personal
Requires a Claude plan with custom connectors (Pro, Max, Team, or Enterprise).
Open claude.ai → Settings → Connectors.
Click the Customize tab to manage connectors.
Name: Marqii. URL: https://mcp.marqii.com/mcp.
Connect → a browser tab opens for Marqii sign-in → close it when it redirects back.
The marqii_* tools now appear in Claude's tool tray.
2. Claude — Team / Enterprise
Org admins install Marqii once; each user still OAuths individually on first use (the Cognito session is per-end-user, not per-org).
Open the Anthropic Admin Console → Connectors.
Add -> Custom -> Web → URL https://mcp.marqii.com/mcp.
Save. The connector is now discoverable for all members under Customize → Connectors.
Each member clicks Connect from Customize -> Connector sthe first time they use it and signs in with their Marqii account.
Org SSO does not flow through to Marqii — users sign in with their Marqii credentials, not their Anthropic SSO identity.
3. ChatGPT (OpenAI)
Available on ChatGPT Pro, Business, Enterprise, and Edu (admins may gate custom connectors).
- Access Settings: Open the ChatGPT web interface, click on your profile name in the left sidebar, and select Settings.
- Enable Developer Mode: Go to the Apps & Connectors (or Connectors) tab, click Advanced, and toggle Developer Mode to On.
- Add the Server: Return to the Connectors menu and click Create.
- Configure Details: Enter a name for your MCP server and paste your remote https://mcp.marqii.com/mcp.
- Configure your authentication type OAuth
4. Ollama
Ollama is a local model runtime; it has no built-in MCP client. The standard pattern is to put an MCP-aware front-end in front of Ollama and point that front-end at mcp.marqii.com. Two practical front-ends:
Continue (VS Code / JetBrains)
Continue speaks MCP and runs against any OpenAI-compatible model server, including Ollama.
Edit ~/.continue/config.yaml:
mcpServers:
- name: marqii
transport:
type: streamable-http
url: https://mcp.marqii.com/mcp
Reload Continue; the first tool call triggers OAuth in your browser.
A CLI bridge (mcp-cli, mcphost, etc.)
Any CLI that supports remote MCP + OAuth can be pointed at https://mcp.marqii.com/mcp and configured to use an Ollama model for inference. Consult the bridge's own README for OAuth handling — most defer to the system browser the first time, then cache tokens on disk.
5. OpenLLM (BentoML)
Same shape as Ollama: OpenLLM is a model server, not an MCP client. Pair it with an MCP-aware front-end (Continue, Claude Code with a custom provider, an mcp-cli bridge, etc.). The front-end calls OpenLLM for completions and mcp.marqii.com/mcp for tools.
Example Continue snippet:
models:
- name: my-openllm
provider: openai
apiBase: http://localhost:3000/v1 # your OpenLLM endpoint
apiKey: dummy
model: <whatever you served>
mcpServers:
- name: marqii
transport:
type: streamable-http
url: https://mcp.marqii.com/mcp
6. Cursor
Edit ~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project):
{
"mcpServers": {
"marqii": {
"url": "https://mcp.marqii.com/mcp"
}
}
}
Reload Cursor (Cmd+Shift+P → Reload Window). Open Settings → MCP to confirm Marqii is listed. The first tool call opens a browser tab for Marqii sign-in.
7. Windsurf
Edit ~/.codeium/windsurf/mcp_config.json (or use Settings → Cascade → MCP servers → Add server):
{
"mcpServers": {
"marqii": {
"serverUrl": "https://mcp.marqii.com/mcp"
}
}
}
Restart Cascade. First call triggers OAuth.
8. Antigravity (Google)
In the Antigravity IDE: Settings → MCP servers → Add server → choose HTTP → URL https://mcp.marqii.com/mcp. Save and reload the agent panel; sign in with your Marqii account when prompted.
If your version of Antigravity uses a JSON settings file instead of a UI:
{
"mcpServers": {
"marqii": {
"transport": "http",
"url": "https://mcp.marqii.com/mcp"
}
}
}
9. Generic MCP client
If your client supports the MCP spec's remote-server profile, give it:
Transport: Streamable HTTP (the spec replacement for SSE; not stdio)
Auth: OAuth 2.1, authorization code + PKCE
Authorization server metadata: https://mcp.marqii.com/.well-known/oauth-authorization-server
Protected resource metadata: https://mcp.marqii.com/.well-known/oauth-protected-resource
Dynamic client registration: POST https://mcp.marqii.com/register (RFC 7591)
Clients that auto-discover the metadata endpoints need nothing else. Clients that require a pre-registered client_id can register one manually:
curl -X POST https://mcp.marqii.com/register \
-H 'Content-Type: application/json' \
-d '{
"client_name": "my-client",
"redirect_uris": ["http://localhost:7777/oauth/callback"],
"token_endpoint_auth_method": "none"
}'
Use the returned client_id in your client's OAuth settings; set the redirect URI to whatever URL your client listens on for the auth callback.
Please note: Marqii MCP is currently only compatible with Gemini CLI. It is not supported on the web or app versions of Gemini at this time.
Troubleshooting
401 / Not Authorized on every call. Your access token has expired and your client did not refresh it. Disconnect and reconnect the connector to re-run OAuth.
"No tools listed" or "406 Not Acceptable". Your client is sending a POST without Accept: application/json, text/event-stream. Almost every maintained client gets this right; if you're rolling your own, fix the Accept header.
You want to sign in as a different Marqii user. Call the marqii_switch_user tool inside the chat — it returns a logout URL that clears the Cognito session, then disconnect and reconnect the connector to OAuth as the new user. See README → Switching users / logging out for the full flow.
Revoke a stored token without re-OAuthing. POST https://mcp.marqii.com/revoke with token=<refresh_token> (RFC 7009). Many clients call this automatically on disconnect.
Connector can't reach the server. mcp.marqii.com is public on the open internet; check your network/VPN. Anthropic, OpenAI, etc. need to reach it from their backend, not just from your laptop.
If you're having issues with setup contact your Account Manager or email support@marqii.com for assistance.
