September 2026 Summaries

LocalStack’s Snowflake emulator, combined with its MCP server and Snowflake client tool, enables AI agents such as Claude Code to build, test, and verify SQL workflows locally without incurring real Snowflake warehouse costs. In the example, an agent loaded SaaS billing CSV data into an emulated Snowflake environment, created a Q2 2026 Revenue by Plan report, and validated its results against raw tables. Initial plausible-looking results contained two common errors: joining payments duplicated customers in active MRR calculations, while filtering for active customers excluded revenue from customers who churned after paying during the quarter. By separating MRR and revenue into distinct calculations and reconciling every value through alternative queries and raw payment totals, the agent produced corrected totals of $5,873 in active MRR, 27 active customers, and $20,741 in collected quarterly revenue. The process took about nine minutes, 42 tool calls, and $2.80 in model usage, illustrating how local emulation and mandatory reconciliation can catch subtle SQL mistakes before reports reach production.

LocalStack enables developers to run Amazon Cognito user pools locally, avoiding shared AWS test accounts, exposed CI passwords, and authentication issues that otherwise may only emerge in production. The tutorial builds an expense-claims portal with Terraform, a React application, Lambda, API Gateway, DynamoDB, and a Node.js payroll script to demonstrate four Cognito patterns: hosted UI authorization-code login with PKCE for employees, JWT validation by an API Gateway authorizer, TOTP-based MFA for finance administrators, and client-credentials access with a custom scope for payroll exports. It explains Cognito tokens, OpenID Connect discovery, JWKS-based token verification, optional MFA configuration, user groups, and route-level scope enforcement, while noting LocalStack differences such as logging confirmation codes and not yet enforcing PKCE verifier validation or hosted-UI MFA challenges. Developers can create isolated pools and users, test sign-up, claims submission and approval, MFA enrollment, machine-to-machine authorization, and rejected user-token access locally, then run an end-to-end integration suite without AWS credentials. The approach provides repeatable infrastructure and faster feedback while preserving Terraform and application compatibility for later AWS deployment.

LocalStack Cloud Pods and its MCP server can help AI agents preserve, inspect, and restore complete local AWS emulator environments, including infrastructure and application data. Using a CDK sample application, Claude Code deployed 36 resources, added 12 DynamoDB records, and saved the working environment as a remote Cloud Pod checkpoint. A seemingly minor DynamoDB table rename then caused CloudFormation to replace and delete the original table because its removal policy was set to DESTROY, leaving the API functional but returning no data. The agent detected the loss through a record-count check, restored the baseline Cloud Pod, reverted the rename, changed the table policy to RETAIN, confirmed that all records remained, and saved a hardened checkpoint. Cloud Pods provide versioned, shareable snapshots that replace current LocalStack state when loaded, enabling teams and CI systems to recover from destructive infrastructure changes without rebuilding environments or re-creating test data.

AWS Lambda Durable Functions, released in late 2025, allow long-running, stateful workflows to remain in a single Lambda handler by checkpointing completed steps and replaying execution after retries, waits, callbacks, crashes, or interruptions. The tutorial demonstrates building an order-processing workflow with Python, CDK, DynamoDB, SNS, and SQS, in which an order is validated and recorded, payment is retried with exponential backoff, an external payment-confirmation callback pauses the workflow, notifications are sent in parallel, and the order is marked fulfilled. LocalStack 2026.8.0 or later supports durable-function creation, execution management, retries, callbacks, parallel branches, history, idempotent execution names, and persistent state, enabling developers to test recovery scenarios locally. The example deploys a production-ready package containing the durable execution SDK, starts a named execution, verifies that failed payment attempts cause separate invocations while completed steps are skipped on replay, restarts LocalStack while the execution awaits a callback, and confirms that persisted checkpoints allow the workflow to resume successfully. After a callback is sent, the execution completes with a fulfilled status, records the charge and confirmation in DynamoDB, and produces email and SMS notifications, illustrating fault-tolerant workflow behavior without a separate Step Functions state machine.

An experiment used an AI agent with LocalStack’s MCP server, Chaos API, and App Inspector tracing to autonomously run a chaos-engineering game day against a Go/CDK serverless pipeline consisting of API Gateway, Lambda, Kinesis, DynamoDB, and CloudWatch. After establishing that 60 normal requests were delivered successfully, the agent injected Kinesis and DynamoDB outages, DynamoDB throttling, network latency, and CloudWatch failures, identifying that short Lambda timeouts produced ambiguous 502 errors, unlimited event-source retries could stall processing, and downstream metric-publishing failures were silently logged rather than reported. It updated the application to constrain SDK retries within Lambda deadlines, return clear 503 responses, process failed batch records selectively, return downstream errors, configure bounded retry policies, split failing batches, and send exhausted failures to a dead-letter queue. Repeated tests showed that the pipeline retained normal delivery, accepted all requests under 800 ms added latency instead of none, preserved records during DynamoDB outages through observable retries, and surfaced CloudWatch failures through invocation errors and DLQ entries. The unattended session took 26 minutes, used $8.23 in tokens, made 37 MCP tool calls, and demonstrated how compatible AI agents can reduce the cost and effort of cloud resilience testing without a production-like AWS staging environment.

LocalStack announced its acquisition of WonderTwin AI in September 2026, bringing together LocalStack’s cloud infrastructure emulation platform and WonderTwin’s local emulators for applications including GitHub, Stripe, PostHog, and HubSpot. WonderTwin founder Tela Andrews argues that as AI agents accelerate software development and introduce dependencies at volumes beyond traditional change-management processes, developers need reliable models of how services actually behave, including rate limits, edge cases, state, and failures, rather than relying only on API specifications. The proposed approach, termed agentic full-stack emulation, aims to ground automated workflows through continuously calibrated local emulation across both third-party applications and cloud infrastructure. WonderTwin currently supports more than two dozen commonly used applications, while LocalStack’s infrastructure emulation is used by over 1,500 organizations and has been downloaded from Docker hundreds of millions of times. Andrews will lead LocalStack’s Application Emulator efforts as the companies work to integrate local application and cloud development emulation into a more comprehensive development environment.

LocalStack is introducing lstk as its recommended unified command-line interface for local development and CI/CD, consolidating environment management, sessions, state, snapshots, AWS interactions, and authentication into one tool. Available through Homebrew or npm, lstk is intended to reduce the operational complexity caused by the existing mix of localstack CLI, awslocal, cdklocal, Python dependencies, and manually managed tokens, while offering browser-based login for developers and CI authentication through the LOCALSTACK_AUTH_TOKEN variable. It also centralizes configuration through TOML files and simplifies state handling with snapshot commands. Beginning September 8, 2026, LocalStack will focus future command-line development on lstk, although the existing localstack CLI will remain functional for the near future and can operate alongside lstk without requiring an immediate migration. Teams are encouraged to adopt lstk gradually by testing local workflows, command mappings, authentication, state management, and CI pipelines before updating shared automation and internal documentation.

AT Protocol (ATproto), created by the Bluesky team, is an open protocol for decentralized social-web data that supports applications beyond short-form social media, including blogging, code hosting, events, and video platforms across an ecosystem called the Atmosphere. Its emphasis on data sovereignty allows users to move accounts and data between compatible providers, while its public firehose enables developers to consume openly available records without OAuth or individual consent. The described demo ingests ATproto Jetstream WebSocket events into AWS, archives raw batches in Amazon S3, normalizes and distributes events through Amazon EventBridge, and routes them to independent Lambda functions for analytics, moderation, and notifications, with DynamoDB cursors and dead-letter queues supporting recovery and duplicate-tolerant processing. It presents two ingestion approaches: an always-on ECS Fargate consumer for near-real-time delivery and a scheduled Lambda poller for lower-cost, higher-latency processing, both tested locally with LocalStack and a mock firehose. The design highlights practical concerns such as unpredictable stream volume, replayed events, finite cursor retention, partial failures, idempotent consumers, archive consistency, monitoring, and the need for backfill mechanisms in production.

LocalStack for Snowflake 2026.08.0 adds compatibility improvements for query status reporting, metadata commands, ordered aggregates, pipes, tags, and result scanning, while updating its Docker base image from Debian sid to Debian trixie. Long-running queries now remain in progress until execution completes, RESULT_SCAN can process output from status-returning commands such as SHOW and DESC, and multiple LISTAGG or ARRAY_AGG expressions can use independent ordering clauses in one SELECT statement. The release also refines schema-scoped SHOW behavior, materialized-view metadata, fully qualified pipe operations, idempotent tag creation, and tag cleanup rules when databases or schemas are removed. Alongside the standard release, LocalStack introduced the invitation-based Snowflake-Next preview, a planned successor emulator that reportedly executes its regression suite more than ten times faster and adds preview support for streams, zero-copy cloning, and Time Travel.

AI coding agents can increasingly build and operate AWS applications, but effective and safe agentic development requires more than detailed prompts: it needs persistent project rules, current AWS knowledge, actionable tools, and a controlled execution environment. The proposed approach combines Cursor Rules or similar rules files to enforce conventions such as Terraform use, local testing, and prohibitions on unapproved production deployments; AWS’s Agent Toolkit, MCP Server, plugins, and skills to provide documentation, authenticated AWS capabilities, and reusable workflows; and LocalStack with its MCP Server to let agents deploy, test, inspect logs, inject faults, and iterate against an AWS-compatible local environment. This local-first model limits the risks and costs of allowing agents to rapidly create or modify real cloud resources while supporting full development loops that include infrastructure creation, testing, debugging, and redeployment. AWS MCP and LocalStack MCP are presented as complementary rather than competing components, with the former supplying AWS knowledge and access while the latter controls the local testing environment. The broader argument is that developers’ roles shift toward defining intent, architecture, constraints, permissions, and release review, while agents handle more of the repetitive implementation and validation cycle.