MCP and AI agents reference

Your Arino One instance exposes a Model Context Protocol (MCP) endpoint at /mcp so an AI client can call a small, scoped set of tools against your workspace, and Infobip separately publishes its own remote MCP servers so an agent can operate its messaging channels directly. These are two different connections serving two different purposes, and this page covers both.

The instance's own MCP endpoint

Every instance runs an MCP handler at /mcp, backed by tools defined in src/lib/mcp/tools. Out of the box this includes a whoami tool (confirms who the agent is acting as) and a list_contacts tool (browses the workspace's contacts). It is deliberately minimal: a way for an AI client to verify connectivity and read basic workspace data, not a general-purpose automation surface.

Connecting an AI client

1. In your MCP-capable client (a chat assistant, IDE agent, or automation tool), add a remote server pointing at https://<your-instance-domain>/mcp.
2. The client discovers the authorization server via the standard metadata path, /.well-known/oauth-protected-resource, published alongside the endpoint.
3. The client initiates an OAuth 2.1 flow against the instance's own auth issuer — the direct Supabase auth host behind your instance, not a proxy domain.
4. Sign in as the user you want the agent to act as. The agent's session then carries that user's permissions.
5. Confirm the connection by invoking whoami; it should return the identity you just authenticated as.

Scopes and approval

Because authentication runs through your normal user session, an agent connected this way sees exactly what that user is allowed to see under row-level security — it cannot read another user's or another workspace's data by virtue of the MCP connection. There is no separate "MCP scope" grid to configure in Settings; access control is inherited from the signed-in identity, which is also why you should only authorise this connection as a user whose data access you are comfortable extending to an automated client.

This is distinct from the Arino Link inbound-command approval flow: any command arriving through Arino Link, however it originated, still passes through the admin Command Inbox before execution. MCP tool calls and Arino Link commands are separate mechanisms, and neither is a shortcut around the other's authorisation model.

Infobip's remote MCP servers

Independently of your instance's own endpoint, Infobip publishes remote MCP servers for operating its channels directly. These are useful while building, testing or maintaining your instance's messaging setup, typically from an AI coding assistant rather than from an end-user chat.

Other channels follow the same URL pattern (https://mcp.infobip.com/{product}) and are documented at Infobip's MCP catalogue.

Connecting

Infobip's servers support Streamable HTTP by default, with legacy Server-Sent Events available by appending /sse to the endpoint (for example https://mcp.infobip.com/sms/sse).

• Preferred: OAuth 2.1. Add the endpoint to your MCP client with no Authorization header. The client initiates the OAuth flow, redirects to Infobip, and manages token refresh itself — no API key is ever pasted into the client configuration.
• Fallback: API-key header, for clients that do not support OAuth 2.1: send Authorization: App <INFOBIP_API_KEY>, using a key scoped for the target server (for example sms:message:send for the SMS server).

The two-credential rule

Keep these two Infobip credentials strictly separate:

1. MCP/agent credential — the OAuth token or API key an agent uses to call Infobip's MCP servers on your behalf while you work.
2. Runtime BYO key — the API key your instance uses at runtime to send real messages to real customers, entered through the setup wizard and stored server-side.

Never paste the runtime key into an MCP client, chat, or configuration file, and never reuse an MCP OAuth token as the runtime key — their scopes and rotation are unrelated. See AI providers and API keys for how the runtime side is stored.

Typical uses

• Verify an Infobip key: connect account-management and call its balance or account-details tool.
• Create a WhatsApp template: connect whatsapp, supply name, language, category and body text, then check the returned status (usually PENDING until Meta approves) — see WhatsApp message templates.
• Send a first test message: prefer the setup wizard's own test-send step so the delivery is logged and auditable through your instance; if sending from chat instead, confirm sender and recipient first, since Infobip trial accounts silently drop sends to unverified numbers.
• Build a WhatsApp Flow: connect whatsapp-flow, iterate on the flow structure, then publish and send a test with the whatsapp server.

Treat any AI agent with MCP access — to your instance or to Infobip — as you would a human with the same credentials: scope it to what it needs, and prefer your instance's own logged, auditable paths (webhooks, the setup wizard, the message API) for anything that touches real customer data.

Frequently asked