POINTS journal · products
Agent Operators
Audience: Developers and power users who put agents on Points rails and own the guardrails.
Positioning: You set the rails; agents run inside them.
Who
Humans (and teams) configuring agents: founders running agent stacks (ChatGPT MCP connector live today; Claude Desktop/Code, Cursor and OpenClaw-class hosts next, no dates), ops who approve agent spend, builders who don’t want to invent a payment policy engine.
Job to be done
Configure allow-lists, caps, counterparties, and mandates; review and approve payment drafts; audit what agents attempted — without building a payment stack.
What Points offers now
- Human approve in Controls / approvals queue
- Policy vocabulary, velocity / anti-fragmentation
- Live host: ChatGPT MCP connector
- Signed envelope + approval ledger recorded alongside each approved draft
- Per-payment and scope checks (principal, counterparties, expiry); aggregate caps (in progress)
- Recurring mandates require a fresh approval each cycle (rolling out)
- Recipient ↔ ACH provider external-account id mapping (no routing/account storage)
- Agents get discoverable constraints; operators stay authoritative
What’s next (no dates)
- Hosts: Claude Desktop/Code, Cursor, OpenClaw-class runtimes
- Operator UI for envelopes, mandates, counterparty allow-lists, aggregate caps
- Discoverable policy editor (not only MCP tools)
- Clear UX: agent may negotiate / human approves
- Mandate schedule jobs (
nextCycleAt) - Audit views: approval ledger vs transaction log
- Webhooks / alerts on refuse, expire, revoke
How it works
Point your agent at POINTS MCP, set policy once, approve payments when they come in. You keep veto; the agent keeps speed inside your bounds.
Operators own risk. Approve ≠ execute until execution is explicitly enabled.