Skip to main content
Control is the rules layer. Two engines — Policy and Risk — co-evaluate every intent, grant, issue, redeem, and checkout. Decisions are server-side; agents cannot bypass their own rules.

​
The two engines

Policy Engine

Pure function evaluatePolicy(policy, input) → { allowed, requiresApproval, reason }. Deterministic, declarative, plan-tier-agnostic.

Risk Engine

Stateful. 5 signals (velocity, amount-anomaly, geo, session, time-anomaly). Score 0–100 with tunable thresholds for review/deny/freeze.

​
Policy: what can be expressed

Policies are JSON-serializable objects attached to an agent (one policy per agent at a time). The full schema lives in /concepts/policies — here’s the surface:

​
Risk: signals and thresholds

Score is a weighted aggregate. Thresholds are tunable per policy (5min cache):
Auto-freeze is opt-in. It will halt legitimate flows if your tolerance isn’t calibrated — recommend turning it on only after observing risk scores in production.

​
Where Control runs

Each evaluation writes a row to decision_logs with the full signal payload, the score, the reason, and the action taken. Queryable via /audit and /risk/violations.

​
Card-controls sync

When you change a policy, the system translates the relevant fields into card-issuer-native controls (MCC rules, country rules, per-tx limits) and pushes them to the underlying card. Policy is the source of truth; the card mirror is best-effort and is reconciled on every update.

​
Approval workflow

When enforcementLevel: "approve" or risk action is review, the intent enters pending_approval. Approve via:
Or POST /intents/:id/deny with an optional reason.

​
Quick example — read an agent’s policy

Or via MCP — agents can read their policy but never modify it:

​
Pre-built presets

policy-presets.ts ships templates for common patterns: travel-only, subscriptions, dev-tools, etc. Use them as starting points in the dashboard create flow.

​
Surfaces

​
Next

Policies

Full field reference with examples for each rule.

Intents

The approval-bearing primitive Control gates.

Intelligence

What Control writes — decision logs, audit, anomaly detection.

Disputes

What happens when something Control allowed turns out to be wrong.