Skip to content
Apiway Docs

Secure my API for AI agents

AI agents are calling APIs. They need the same governance as human consumers — identity, scoped permissions, rate limits, and audit trails.

What you need

Section titled “What you need”
  • A deployed API on Apiway (ship one first if you haven’t)
  • An AI agent that supports MCP (Claude Code, Copilot, or your own)

The pipeline

Section titled “The pipeline”

  1. Your API is already MCP-ready

    Every API deployed through Apiway with a well-described OpenAPI spec is automatically discoverable via MCP. The gateway exposes your endpoints as tools that AI agents can find and call.

    No extra configuration. If your spec has good summaries and parameter descriptions, it works.

  2. Create an agent identity

    Don’t share human credentials with AI agents. Create a dedicated identity:

    Terminal window
    curl -X POST https://core.api.apiway.net/v1/subscriptions \
      -H "Authorization: Bearer $TOKEN" \
      -H "Content-Type: application/json" \
      -d '{
        "apiId": "{api-id}",
        "consumer": "checkout-agent",
        "type": "agent"
      }'

    The agent gets its own credentials, its own scopes, and its own audit trail — separate from every human user.

  3. Scope the permissions

    Agents shouldn’t have access to everything. Restrict to specific operations:

    Terminal window
    curl -X PUT https://core.api.apiway.net/v1/subscriptions/{id}/scopes \
      -H "Authorization: Bearer $TOKEN" \
      -H "Content-Type: application/json" \
      -d '{
        "scopes": ["<scope to read orders>", "<scope to create orders>"]
      }'

    Each operation’s scope is shown with it in your API’s entry in the portal. The gateway enforces this. If the agent tries to call DELETE /orders, it gets a 403. No trust required.

  4. Set rate limits and cost guards

    Protect against runaway agents:

    Terminal window
    curl -X PUT https://core.api.apiway.net/v1/subscriptions/{id}/limits \
      -H "Authorization: Bearer $TOKEN" \
      -H "Content-Type: application/json" \
      -d '{
        "rateLimit": 500,
        "unit": "minute",
        "budgetCeiling": 1000,
        "budgetUnit": "GBP"
      }'

    Soft alert at threshold. Hard stop at ceiling. The agent can’t run up an uncontrolled bill.

  5. Connect your AI assistant

    Point your MCP-compatible agent to the gateway:

    {
      "mcpServers": {
        "apiway": {
          "url": "https://alpha.gateway.apiway.net/mcp",
          "headers": {
            "Authorization": "Bearer $AGENT_TOKEN"
          }
        }
      }
    }

    The agent discovers available tools, understands parameters, and makes governed calls — all through natural conversation.

  6. Audit everything

    What an agent was entitled to do, who approved it, and what it consumed:

    Terminal window
    curl https://core.api.apiway.net/v1/subscriptions/{id}/operations \
      -H "Authorization: Bearer $TOKEN"

    The agent’s entitlements, the approvals behind them, and its consumption per period are resources you query like any other — see Access Evidence. Refused requests are recorded with the rule that refused them, the operation and the source.

    Request-level logging of every successful call, with parameters and responses, is not held by the platform. On Hybrid or Self-hosted the gateway runs on your own infrastructure, so that goes to whatever you already collect logs with. Settle it early if your regulator requires it.

What you got

Section titled “What you got”
  • Agent identity — dedicated credentials, not shared service accounts
  • Scoped permissions — per-operation access control enforced at the gateway
  • Cost guards — budget ceiling prevents runaway consumption
  • MCP discovery — AI agents find and use your APIs automatically
  • Evidence on demand — what the agent may do, who approved it, what it consumed, and every refusal

Next steps

Section titled “Next steps”