Stripe Restricted Keys

Stripe restricted API key permissions — full map

Every permission on the Restricted Key form, what it unlocks, what breaks if you leave it at None, and how much money is exposed if an agent loops on that endpoint.

TL;DR

Restricted Key permissions are grouped by Stripe resource (Charges, Refunds, Customers, etc.). Each resource offers None / Read / Write. Write implies Read. The rows below are the ones most agents need; the runaway-risk column is what matters when the agent gets stuck — not what the feature does under normal use.

How to read the table

Three columns you should weigh before ticking any permission:

What it enables. The concrete agent action — not the docs description, the thing the agent will actually do on a Tuesday at 2 a.m.
What breaks if None. Which agent workflows stop working. Useful to decide whether to tick this for a specific use case.
Runaway risk. If the agent gets stuck calling this endpoint in a loop, what's the worst case by morning? Measured in money moved, customers affected, or external systems pinged.

Core payments permissions

Permission	What Write enables	What breaks if None	Runaway risk
Charges	Create, update, capture charges	Cannot create new charges or capture pre-authed ones	High — loop creates real charges against saved cards
Refunds	Issue refunds on existing charges	Cannot refund	High — loop refunds entire recent charge history
PaymentIntents	Create, confirm, cancel payment intents	Cannot use the modern payments API at all	High — same risk profile as Charges; the modern path
SetupIntents	Attach payment methods for future off-session use	Cannot save cards for later	Medium — no money moves, but enables future charges
Disputes	Submit evidence, accept liability, close disputes	Cannot interact with disputes	Medium — can unnecessarily accept chargebacks

Customer and PM permissions

Permission	What Write enables	What breaks if None	Runaway risk
Customers	Create, update, delete customers	Cannot manage customer records	Medium — can corrupt customer data; combined with Charges, unbounded
PaymentMethods	Attach, detach, update payment methods	Cannot move cards between customers	Medium — card-hygiene disasters more than money loss
Sources / Cards / Bank Accounts	Legacy payment-method surfaces	Legacy integrations break	Low — most agents should leave these off entirely
Tax IDs	Attach tax IDs to customers	Cannot set customer tax IDs via API	Low — data quality issue, not financial

Subscription & invoicing permissions

Permission	What Write enables	What breaks if None	Runaway risk
Subscriptions	Create, update, cancel subscriptions	Cannot manage recurring billing	High — can mass-cancel or mass-upgrade the book
Invoices	Draft, finalize, pay, void invoices	Cannot issue invoices	High — finalizing a looping draft charges the customer
Invoice Items	Add line items to upcoming invoices	Cannot pre-seed invoice lines	Medium — feeds Invoices risk if combined
Credit Notes	Issue credit notes against paid invoices	Cannot issue credits	Medium — unbounded credits are real money out
Plans / Prices	Create and update pricing objects	Cannot programmatically manage catalog	Low — new objects don't auto-bill anyone
Products	Create and update products	Cannot manage catalog	Low — same as Plans
Coupons / Promotion Codes	Create discount artifacts	Cannot programmatically issue discounts	Medium — loop creates uncapped discount codes

Payouts, Transfers, Balance

Permission	What Write enables	What breaks if None	Runaway risk
Payouts	Create manual payouts to your bank	Cannot trigger manual payouts (auto-payouts unaffected)	Very high — moves funds out of Stripe
Transfers	Move funds to connected accounts	Cannot split funds in Connect flows	Very high — Connect platforms should think twice
Balance	Read-only on the balance endpoint	Cannot read available balance	None — read-only
Balance Transactions	Read the ledger	Cannot reconcile	None — read-only

Connect-specific permissions

Permission	What Write enables	What breaks if None	Runaway risk
Accounts	Create, update, delete connected accounts	Cannot onboard new merchants	High — can orphan or corrupt merchant data
Account Links	Generate onboarding URLs	Cannot redirect merchants to the Stripe-hosted flow	Low — just URLs
Application Fees	Refund platform fees	Cannot return platform fees	Medium — mass-refund of fees = mass revenue loss
Capabilities	Request capabilities on connected accounts	Cannot programmatically request card payments etc.	Low — administrative

Supporting and metadata permissions

Permission	What Write enables	What breaks if None	Runaway risk
Webhook Endpoints	Create, update, delete webhook subscriptions	Cannot programmatically wire new webhooks	High — can reroute webhooks to an attacker-controlled URL
Events	Read event history	Cannot paginate audit data	None — read-only
Files	Upload files for disputes / identity	Cannot upload evidence	Low — storage only
Reports	Trigger report runs	Cannot fetch large reports	Low — reporting only
Radar Sessions / Reviews	Manage fraud signals and manual reviews	Cannot close Radar reviews programmatically	Medium — can approve fraud if paired with Charges
Sigma / Data Exports	Run Sigma queries	Cannot run Sigma programmatically	Low — read-only analytics

Permissions most agents should leave at None

A useful heuristic: if your agent's job description doesn't literally reference the feature, leave the permission at None. In particular, we recommend default-off on Payouts, Transfers, Webhook Endpoints, Subscriptions, and Invoices unless the agent explicitly manages those surfaces. The bar to enable is "this agent's intended behavior requires this resource", not "this agent might find this useful."

Where the Restricted Key model runs out

Three classes of problem Restricted Keys cannot solve, no matter how carefully you pick the permissions:

Amount limits. Refunds: Write is "any refund, any amount, any count." No dollar ceiling. A stuck loop exits only when Stripe's rate limit does.
Record-level scoping. Charges: Write is all charges, not just the ones the agent created. You cannot restrict the key to a cohort.
Mid-run revocation. Rotating a Restricted Key takes human time in the Dashboard. If the loop is running, you have a 60-90 second exposure window while a human logs in, finds the key, and clicks revoke.

Read the longer answer on which Stripe-native controls cover which agent scenarios, and where you still have to build your own.

How Keybrake helps

Keybrake attaches three things the Restricted Key itself can't: a daily dollar cap per resource, a customer / merchant allowlist, and sub-second revocation that propagates to every in-flight agent on the next call. You keep the Restricted Key as our upstream; your agents get Keybrake-issued vault keys with the additional policies layered on top. See the full feature set →

Get early access