New purchases: prepaid API budget €10–€10,000 including applicable taxes.

One confirmed payment. No subscription or automatic recharge. Invoice reconciliation (€0.005 HT), supplied OpenAPI contract check (€0.010 HT) and event batches of 1–200 (€0.002 HT) share the net balance. Other suite APIs are not included in current prepaid access.

Choose your budget · Use the three available APIs

Agent Action is an existing control interface, not one of the three prepaid operations. It does not execute or pay for a tool. New integrations require a separate scope review.

Owner-controlled policy · atomic budgets · beta

Set the rules before an agent asks to act.

Keep policy administration separate from the agent's API key. Bind a proposed action to its audience, scopes, exact parameters and declared amount. Reserve the budget, then commit or cancel it explicitly.

Separate authority

A private management key configures policies and issues confirmations. The execution agent receives only its tenant API key.

Bounded spending

Atomic monthly reservations count across policy versions. Concurrent requests cannot exceed the configured budget.

Traceable controls

Request fingerprints, single-use approvals and signed reports expose the control decision. No external action or payment is executed.

Five precise endpoints

https://pfxxzsnboehxsefyajml.supabase.co/functions/v1/xynthera-ai-agent-action
POST /v1/policies
POST /v1/approvals
POST /v1/actions/reserve
POST /v1/actions/commit
POST /v1/actions/cancel

Every call requires x-api-key and application/json. Policy and approval administration also require x-management-key. Bodies are limited to 32 KiB. Every admitted operation, including a business DENY, uses one shared Workbench check. Existing subscriptions retain their accepted combined allowance. No automatic paid overages.

Start with the owner, then connect the agent

Generate and save a private management key locally: 32 random bytes as 64 lowercase hexadecimal characters.
Configure a server policy before distributing the tenant API key to an execution agent. Send both keys for administration.
For a confirmation-required action, request an owner approval bound to its precise action fingerprint. The opaque token is returned once.
Reserve with the execution key and any required approval token. Inspect the signed result; HTTP 200 can contain DENY.
Commit the reservation only through your own trusted integration, or cancel before commit. Commit records budget consumption; it does not run a tool.

The first bootstrap is authorised by the tenant API key plus a locally generated management key. There is no granular owner enrolment or automatic management-key recovery in this beta. Do the bootstrap before sharing agent credentials and keep your management key private.

What the evidence means

The parameters and proposed amount are supplied by your integration. A separate executor must enforce them against the actual tool request. Owner confirmation is not proof of independent human approval. No bank transaction, provider charge or external execution is attested.

Policies allow exact agent, audience, action and scopes, without wildcards. At most 20 rules and 20 scopes per rule; reservation and confirmation lifetimes 1–300 seconds. Monthly declared budget up to 100,000,000 minor units. The policy identifier is fixed; changes require a new version and cannot reset committed monthly consumption.

Transport request_id retries bind exact JSON bytes; reuse the original serialised body. Semantic action input_hash is canonical and separate. Approval tokens are outside signed reports and persisted caches. For existing subscription access, reports are retained under the accepted terms; the budget ledger covers the current and previous two UTC months. Original action parameters are not stored.

Request an Agent Action integration review Read the exact API contract

Full integration guide · Generate a current example policy · Combine with invoice consistency checks