The suite · Reports + Sentinel
Your agents answer for themselves now.
Until now, most operators only heard from their firewall at the moment something was denied. These tools are the other half: ask, in plain language, inside whatever agent you already use, and get answers computed from the tamper-evident audit chain. Read-only, works offline, never gated, and your data never leaves your machine.
Four questions, four tools
“What did my agents spend this week?”
spend_summary
Paid, blocked, by which rule, to which payees, over any window. Totals per currency, honest floors when older records predate amount logging.
“What got blocked?”
list_decisions
Every decision, newest first: payee, amount, the agent's own stated reason at the time, and for denials the exact rule that stopped it.
“Why was this one denied?”
explain_decision
Plain language, not a rule name: what was requested, what the mandate did, why, and whether the audit chain is intact.
“Anything unusual?”
sentinel_alerts
The predictive sweep: which decisions sat outside this agent's own pattern, each alert naming its rule and its numbers.
try them without installing anything: the console on the home page runs these exact functions over a demo week
Sentinel: the warning arrives before the money moves
The mandate answers “is this allowed”. Sentinel answers the question that arrives one step later in every real incident: it was allowed, but is it NORMAL for this agent? Five deterministic signals, computed from the agent's own history at decision time, delivered inside the ALLOW itself, the last moment a human can still act cheaply. No model, no score, no black box: every alert names its rule and its numbers.
Two minutes, any agent
One MCP entry gives Claude, Cursor, or any MCP host the whole surface: the payment firewall, the reports, and Sentinel. OpenClaw agents get the same tools from the native plugin. Same functions, same answers, whichever shell your agent runs in.
{
"mcpServers": {
"fidacy": { "command": "npx", "args": ["-y", "@fidacy/mcp"] }
}
}