AI Agents

Audience: Autonomous agents (and their builders) that need to pay for tools, APIs, inference, and micropayments under policy.

Positioning: Constrained commerce, not a wallet with legs.

Who

Self-driven agents. Live today through the ChatGPT MCP connector; Claude Desktop/Code, Cursor and OpenClaw-class open hosts are next (no dates). They negotiate intent; they do not own the keys.

Job to be done

Plan and prepare payments, discover what policy allows, and receive a human-approved signed authorization — without bypassing the operator’s rails.

What Points offers now

MCP prepare_payment → policy evaluate → human approve → signed spend envelope (executed: false)
The human approves the draft; a signed envelope and an approval ledger entry are recorded alongside it
Scoped approval (limited): principal, counterparties, expiry and per-payment caps are checked; unattendedAllowed defaults false; aggregate caps (in progress)
Recurring mandates: recurring mandates require a fresh approval each cycle (rolling out)
Reservation + cash buckets (unreserved / reserved / escrowed); atomic reserve→settle scaffolding; policy schema versioning; revoke + read-time expiry

What’s next (in development, no dates)

Settlement preflight reasons · independent pre-execution check · plan-time vocab enum + canonical schema · rejection accepted[] · machine-readable receipts · collateral-bounded self-approval · small, closely watched execution trials once spend is enabled

How it works

Connect ChatGPT MCP → prepare under policy → operator approves the draft → agent gets machine-checkable authorization. Spend stays gated while execution is not enabled.

Approve ≠ execute. Autonomy is graduated and operator-gated.

Back to the journal