Policy engine // live
Your agent can pay.
Should it?
MandateGuard in brief — what it is, who it's for, what it costs
What it is. MandateGuard is a spending-policy engine for AI agents. An agent's mandate is a signed, scoped permission — amount caps, allowed vendors, approval rules — and every spend request is checked against it before money moves.
Who it's for. Teams whose AI agents can purchase, book or subscribe, and who need a hard limit plus an audit trail instead of trust.
What it costs. Free covers 1 agent and 1 policy with the full authorize flow and receipts. Pro is $29/month and Team is $79/month. See pricing.
How it works. The agent calls the authorize API (or the MCP tools); MandateGuard answers ALLOW, DENY or REQUIRE_APPROVAL, writes a hash-chained receipt, and a human approves from the inbox when the mandate says so. A kill-switch freezes the agent instantly.
Why now. Agent payment protocols (such as Google's Agent Payments Protocol, announced September 2025) are built around mandates — verifiable, scoped spending permissions. MandateGuard is the policy engine that issues, enforces and audits them.
MandateGuard is built by Ingenious AI.
AP2, UCP and x402 prove a payment was authorised. They don't decide whether this agent, on this task, should spend this money with this vendor again. MandateGuard is that decision — enforced before the spend, stamped into a tamper-evident receipt chain.
▸ Run the live demo See pricing
$12.40
ALLOW
$240.00
REQUIRE_APPROVAL
$89.99
DENY · not allowlisted
Illustrative chain format — the demo below writes real receipts into a real chain you can verify.
▸ Live demo — the real API, one click
Creates a real workspace, agent and policy (per-task cap $50, per-day cap $100, approve over $25, allowlist: openai, aws), then fires four authorize calls at the live engine.
▸ How it works
01 · REGISTER
Register an agent, set a policy: per-task cap, per-day cap, vendor allowlist, approve-over-$X threshold.
02 · AUTHORIZE
Before spending, the agent calls POST /authorize. The engine answers ALLOW, DENY (with reason), or REQUIRE_APPROVAL — in milliseconds.
03 · APPROVE
Approvals land in your web inbox (and your webhook, if configured). Approving releases a one-time spend token the agent redeems exactly once.
04 · PROVE
Every decision is a signed, hash-chained mandate receipt. Recompute the chain any time; export the evidence pack for auditors and insurers.
▸ Paid depth
VENDOR TRUST, FROM YOUR OWN LEDGER
Trust scores computed only from your workspace's receipt history — allowed/denied counts shown beside every score. No invented data, ever.
DUPLICATE-COMMITMENT DETECTION
Same vendor + same amount authorised inside your window? The second one is flagged to a human before money moves twice.
ANOMALY KILL-SWITCH
Freeze an agent instantly. While frozen, every authorize call is refused — stamped DENY · agent_frozen in the chain.
MCP SERVER
Agents check a policy decision and fetch receipts over MCP (POST /mcp, tools mandate.preview + mandate.get_receipt) — the policy layer, native to the agent stack.
INTEROPERABLE INTENT SERVICE
An AP2-style mandate, aligned with the EMVCo draft framework: register an intent once — principal, agent, merchant scope, cumulative + per-period budgets, recurring authority window — and agent, merchant and issuer-side checkers all reference the same intent ID, the same budget, and the same hash-chained event log, refunds and adjustments included. Role-scoped keys: merchants verify, they never write. Open the registry →
POST /authorize
{"agent_key":"mgag_…","vendor":"openai","amount":12.40,"task_ref":"ticket-881"}
→ {"decision":"ALLOW","reason":"within_policy",
"receipt":{"seq":41,"hash":"a91f…","signature":"…"}}