Give your agent the firewall.

The five-minute path to wiring Fidacy into an agent runtime. If you already call HTTP APIs, use the standard Quickstart. This one is for agents that reach for tools — MCP-aware runtimes that call the firewall themselves.

1 · Add the MCP server

Point your runtime at the hosted MCP endpoint. One entry, and the fidacy.assess tool appears in the agent's toolbox.

{ "mcpServers": {
  "fidacy": {
    "url": "https://mcp.fidacy.com",
    "headers": { "Authorization": "Bearer fky_test_…" }
  }
}}

Prefer to run it yourself? See MCP and Hosted MCP.

2 · Teach the agent when to call it

One line in the system prompt is enough. The agent calls the firewall only when a payment is on the table — not on every turn.

# system
Before executing any payment, call
fidacy.assess with the mandate. Only proceed
on "approve"; on "review" ask for confirmation;
on "deny" stop and explain.

3 · Handle the verdict

The tool returns structured output the model can reason over — decision, risk, reasons — plus the signed riskPayloadJws to carry downstream.

→ fidacy.assess
{ "decision": "approve", "risk": 0.12,
  "reasons": [ "mandate-valid", "agent-known" ],
  "riskPayloadJws": "eyJ…" }

4 · Test every branch

Your fky_test_ key runs the real engine against sandbox data. Use magic mandate values to force approve, review, and deny before you go live.

That's the whole loop

Add the server, prompt the agent, branch on the verdict. Explore the capabilities behind it, or try it live with no key at all.