Stripe describes claimable sandboxes, anonymous and ephemeral test environments that let AI agents and partner platforms provision working Stripe API keys without signup. The post covers the design of a hot pool of pre-provisioned sandboxes on Temporal, TTL-based cleanup for cost control, and a terminal-based proof-of-work challenge that protects the endpoint without browser captchas. It also explains how claiming works through account reparenting.
Stripe engineering blog
Loading
Stripe's guide to choosing among its app extensibility primitives: full-page apps, scripts, remote functions, custom objects, and Stripe Workflows. It compares metadata with custom objects and contrasts script and remote-function constraints, including the 30-second and 20-second timeouts. It also stresses signature verification and idempotency, since Stripe retries timeouts and 5xx responses.
Stripe's post explains two ways AI agents can access a Stripe account: acting on a user's behalf through the MCP server with human confirmation for live-mode refunds and payouts, or as independent actors using agent-tagged API keys. It details Approvals rules, the approval_required error flow, reviewer decisions, and the webhook events agents can use to react to outcomes.
Stripe describes how to start a Stripe Atlas company formation application from an AI agent via its MCP server or from the Stripe CLI tools plugin. The post walks through collecting and validating company and founder details, then handing sensitive steps off to the Dashboard, with application state shared across all three interfaces.
Stripe describes how it scaled Shield, the transaction fraud model in Radar, from 800 million to 11 billion training transactions by cutting epoch time from 24 to 4.8 hours through data loading, compilation, and communication tuning. The post also covers a label-maturity problem that degraded recall when fraud reports arrived late, and an agent-driven experiment workflow that improved recall by over 8%.
Stripe announces recurring schedule triggers in Stripe Workflows, which start a workflow at set times without a self-hosted cron job. The post compares this with a typical cron-and-Slack script, covering the operational costs of servers, secrets, and daylight saving time. It walks through a past-due subscription alert that uses a per-invoice metadata marker to avoid duplicate notifications, and explains that a duplicate message is preferred over a missed one.
Stripe explains how labels-only supervised fine-tuning adapts a 4B-parameter LLM for high-stakes risk assessment using serialized heterogeneous business data. The post details probability-based risk scoring, precision-recall results, and probing evidence that fine-tuning improves recognition of legitimate businesses and reduces false positives.
Stripe demonstrates how to build a whitelabeled connected-account dashboard with embedded payments, payouts, and notification components. The walkthrough covers account-session setup, role-based feature permissions, theming, localization, and React integration using Stripe Connect.
Stripe describes how it built an agentic factory for repeated payment-method integrations using more than 100 reusable prompts, observer agents, coded orchestration, and autonomous testing. The approach reduced integration timelines from up to six months to two-to-six weeks and cut one benchmark task from roughly 20 days to four.
Stripe explains how it integrated WebMCP into Checkout to give browser-based AI agents state-aware tools instead of requiring DOM interpretation. Progressive tool disclosure and shared form logic reduced token consumption by 42%, tool calls by 38%, and checkout time by 39% in tests across six models.
Stripe demonstrates how to replace SFTP-based financial reconciliation with flexible, event-driven reporting workflows using the Reports API v2 and Query Run API. The post provides SQL patterns for multi-account payout reconciliation, foreign exchange reporting, and itemized IC+ fee analysis.
Stripe explains how Harbor enables AI-assisted, browser-based prototyping with React and TypeScript. The post details in-browser compilation, agent-driven inspection, resilient comment anchoring across rewrites, and integrations through MCP and the reusable Drydock rendering engine.
Stripe explains how Scripts, Workflows, Custom Objects, and UI extensions let businesses customize logic, data models, and Dashboard experiences within Stripe. Using a vehicle-rental example, the post shows how these capabilities can replace surrounding orchestration infrastructure while coordinating business-specific operations.
Mike North explains how AI reshaped Stripe’s extensibility SDK workflow beyond code generation. The team used rapid prototypes, agent-drafted specifications, repository-local guidance, and faster feedback loops to evaluate architecture earlier and produce more consistent implementations.
Stripe explains why it replaced Envoy with mesh-proxy, a Go-based service-mesh data plane designed to reduce connection cardinality and improve reliability at massive scale. The post details its concurrency model, service discovery, routing features, and reported results of roughly half the CPU usage and a 50% latency reduction.
The post explains how AI-generated payment integrations can be technically correct while violating unstated business rules. It shows how to detect signals such as stuck PaymentIntents, unsafe retries, orphaned authorizations, and mishandled asynchronous payments, then encode those rules in metadata, verification agents, and CI checks.
Stripe explains how Connect platforms can move from blended pricing to IC++ network cost passthrough, including eligibility requirements, Dashboard and API setup, balance recovery behavior, and reporting options. It also outlines a pilot-based rollout strategy for connected accounts.
Stripe presents an architecture for letting autonomous agents shop on behalf of households without exposing reusable card credentials. Using ACP, hosted marketplace surfaces, deterministic household policies, bounded payment authority, SPTs, and idempotent order reconciliation, the smart-fridge example shows how to keep agentic purchases safe and auditable.
Stripe explains how it built Kai, a surface-agnostic knowledge AI platform for secure, multi-turn work across more than 1,000 internal tools and skills. The post covers its APIs, AgentStudio control plane, per-session sandboxes, context isolation, orchestration, and adoption results across the company.
Stripe describes an auto-remediation system that models MongoDB shard infrastructure as a state graph and uses BFS and Dijkstra search to generate safe recovery plans. An in-memory simulation layer shared with Temporal execution preserves fidelity while reducing pager volume by 30% and enabling support for new shard layouts without manual workflows.