THE SPENDING GATE FOR AI AGENTS

Before your agent spends,
make it ask permission.

One preflight checks the owner’s mandate, vendor trust, required evidence and budget—then returns a clear pass, review or block receipt.

Run a live decisionConnect in 60 seconds
NO CUSTODYNO WALLET ACCESSNO PAYMENT EXECUTIONSTATELESS SANDBOX
LIVE POLICY PREFLIGHTPUBLIC ENDPOINT
NO PAYMENT EXECUTED
AGENT INTENTPurchase a data serviceOwner mandate · 100 per request · 250 remaining today

Calls the live stateless endpoint. The request is not stored; Agent Reserve holds no funds, accesses no wallet and executes no payment.

<1minute to first live receipt
3deterministic outcomes
0funds or credentials stored

Public developer sandbox. Demand and production integration remain unproven.

ONE QUESTION · THREE SAFE OUTCOMES

Should this agent request
be allowed to continue?

Agent Reserve gives downstream systems a small, predictable decision surface. Uncertainty never silently becomes a payment.

PASS

Known vendor. Valid evidence. Inside budget.

The agent may continue to the owner’s existing payment workflow.

REVIEW

Unknown vendor, missing proof or budget exception.

The request pauses and returns to a human before any payment rail is invoked.

BLOCK

Invalid mandate, blocked vendor or failed evidence.

The request stops. No payment workflow should be invoked.

BUILT FOR AGENT DEVELOPERS

Put policy between intent and payment.

Use the same preflight before any irreversible step. Agent Reserve never replaces the owner or the payment rail—it makes the boundary explicit and auditable.

01

API and data purchases

Let research agents buy approved datasets while stopping unknown vendors and missing delivery evidence.

02

Cloud and compute

Keep autonomous workloads inside per-request and daily spend limits before usage is provisioned.

03

SaaS procurement

Require an owner mandate and a trusted vendor before an agent starts or renews a software subscription.

MCP + OPENAPI · PUBLIC SANDBOX

Connect in 60 seconds.

Give any compatible agent one read-only tool. It sends six bounded facts and receives a deterministic decision receipt. No account, wallet or secret is required for the sandbox.

ENDPOINT STATELIVEStateless · public developer sandbox
  1. 1
    Add the MCP endpointhttps://agent-reserve-protocol.super-myna-7255.chatgpt.site/api/mcp
  2. 2
    Call the spending gateevaluate_agent_spend_request

    Pass the mandate, vendor, evidence and budget facts before an irreversible action.

  3. 3
    Honor the result

    Continue on eligible, pause on escalate, and stop on block or no valid response. Eligible never executes payment.

EXAMPLE REQUESTREST /api/authorize
{
  "mandateValid": true,
  "vendorStatus": "trusted",
  "receiptStatus": "pass",
  "amount": 48,
  "perRequestLimit": 100,
  "dailyRemaining": 250
}
RETURNSdecision + reason + SHA-256 receipt

COPY · CONNECT · VERIFY

Two-minute activation test.

ACCEPTANCE: ONE VALID RECEIPT
01 · CLIENT CONFIG

For MCP clients that accept remote Streamable HTTP server entries.

{
  "mcpServers": {
    "agent-reserve": {
      "type": "http",
      "url": "https://agent-reserve-protocol.super-myna-7255.chatgpt.site/api/mcp"
    }
  }
}
02 · FIRST AGENT CALL

Paste this instruction after the server appears in your client.

Before any payment step, call evaluate_agent_spend_request with:
mandateValid=true, vendorStatus=trusted, receiptStatus=pass,
amount=48, perRequestLimit=100, dailyRemaining=250.
Continue only on eligible. Stop on escalate, block, or no valid response.
03 · EXPECTED RECEIPT

The integration passes only when the tool returns all four facts.

  • decisioneligible
  • codepolicy_pass
  • paymentExecutedfalse
  • requestFingerprint64 hex chars
Fail closed: if the tool is unavailable, malformed, or returns anything other than eligible, do not invoke a payment rail.
MCP endpoint ↗Server manifest ↗OpenAPI 3.1 ↗8 test vectors ↗

Integration is not enforcement by itself. The downstream system must stop when Agent Reserve returns review, block or no valid response.

TRUST THROUGH RESTRAINT

It cannot move money.
That is the point.

01Owner policy stays primary

The agent supplies policy facts; the owner defines the actual mandate and limits.

02Every result is reproducible

Exact validation order and deterministic decision vectors make results inspectable.

03Every receipt is bounded

A policy pass is not a transaction, legal approval or wallet instruction.

04Failures stop safely

Invalid requests and failed evidence block instead of falling through.

OPEN PROOF ARCHIVE

Inspect the policy, tests and safety boundaries.

The former protocol laboratory remains available as machine-readable evidence. It is proof of deterministic reference logic—not certification, customer demand or production security.

START WITH ONE REQUEST

Give your agent a spending boundary before it needs one.

Run live preflightSandbox only · No account · No payment · No stored request