MCP

Connect an AI agent to Rivier over the Model Context Protocol with one address and a key.

Connect

The server is at https://api.rivier.ai/mcp. It speaks Streamable HTTP and keeps no session, so one address and one key are all an agent needs. Send the key as a bearer credential, the same as REST. An agent that signs in with OAuth finds the issuer at /.well-known/oauth-protected-resource/mcp.

Configuration:

{
  "mcpServers": {
    "rivier": {
      "type": "http",
      "url": "https://api.rivier.ai/mcp",
      "headers": { "Authorization": "Bearer ${RIVIER_API_KEY}" }
    }
  }
}

Tools

The tool list follows the scopes of the key: a read key sees read tools, a write key also sees the writes its scopes allow. Three tools reach every operation in the API reference, so nothing in the app is missing for an agent.

• rivier_api_operations: Find an operation by words, area or level. Answers the id, the route, the level and a summary.
• rivier_api_read: Run a read operation with its parameters and query.
• rivier_api_act: Run a write operation with its parameters, query and body. Takes idempotencyKey so a retried write is not repeated.
• rivier_key_scopes: What the calling key holds: project, scopes and how many operations of each level it can run.
• rivier_overview: Contacts, campaigns and flows in one call.
• rivier_contact_brief: One person: the record, recent events and consent history in one call.

Resources and prompts

Resources: rivier://guide, rivier://scopes and rivier://operations. Prompts: weekly_digest, contact_follow_up and draft_campaign_brief. None of the prompts sends or approves anything.

What an agent cannot do

A call goes through the same route as REST with the same key, so scopes, project and rate limits are unchanged. Actions kept for a person, such as approvals, sends and removals at the admin level, are refused with the reason, and the agent can leave them for a person to approve in the app.

When a call fails

A failed call answers a fixed shape. Branch on code, not on the wording. hint names the next step, param the fields at fault, and next the tool and arguments to call instead where there is one. Long answers are cut at a record boundary with a note to narrow the query; nothing is cut without saying so.

Response:

{
  "ok": false,
  "code": "insufficient_scope",
  "retryable": false,
  "error": "…",
  "hint": "…"
}

Codes: invalid_input, insufficient_scope, not_found, conflict, rate_limited, missing_params, unknown_operation, wrong_mode, reserved_for_person and server_error.