Twelve new services that really run, and 80% of AWS's operations. EFS, WAF, Timestream, IAM Identity Center, CodeCommit, Managed Flink, DMS, EMR on EC2 and more; pipelines that run, EC2 disks with real data, live RDS replication and Airflow 3. See what's new.
Mimir is now Goku, everywhere. Every name is Goku now: the image tanujsoni027/goku, the goku container and command, and GOKU_* settings. Existing installs move over automatically and the old names keep working for now, so start using the goku names.
Goku (formerly Mimir) is a local AWS cloud emulator: 87 AWS-compatible services — S3, Lambda, DynamoDB, RDS, Step Functions, API Gateway, Glue and more — on localhost:4566, with a web console for 70 services on localhost:8080, real container-backed engines, and zero AWS account. Free.
Install Goku What's new in 4.3$ curl -fsSL https://tanuj24.github.io/goku/install.sh | sh
Then run goku start and open http://localhost:8080 — click “Load sample data” on any console page to explore a populated cloud.
Or run it with Docker directly — see all install options. Needs Docker Desktop on macOS and Windows, or Docker Engine or Docker Desktop on Linux.
Goku 4.3 is the biggest release yet. Twelve new services run for real, from Amazon EFS and AWS WAF to Managed Service for Apache Flink, DMS and EMR on EC2. Pipelines run their stages, EC2 disks hold real data, RDS replicas follow their source live, and MWAA runs Airflow 3 with Celery workers. Goku now answers 5,162 of 6,441 AWS operations (80%) across 87 services.
Amazon EFS file systems with real files, AWS WAF web ACLs that filter real traffic, Amazon Timestream with its SQL, IAM Identity Center with a working access portal, IAM Access Analyzer, AWS CodeCommit Git repositories you clone and push to, Managed Service for Apache Flink, AWS DMS, Amazon EMR on EC2, MWAA Serverless, CloudFront KeyValueStore and AppSync Event APIs.
DMS tasks copy MySQL, MariaDB, PostgreSQL and S3 data into databases, S3, Kinesis, DynamoDB or OpenSearch, then keep replicating changes. Flink applications run your jar on a real Flink cluster. EMR clusters run Spark steps on real nodes. RDS read replicas and blue/green deployments of MySQL, MariaDB and PostgreSQL replicate live, and a switchover loses no write.
CodePipeline runs its stages on Goku's own services: a push to CodeCommit or ECR, or a new S3 object, starts it, CodeBuild builds in real containers, and CloudFormation, ECS, CodeDeploy and S3 deploy, with manual approvals and Lambda and Step Functions actions. MWAA runs the Airflow versions MWAA offers today, Airflow 3 included, with Celery workers that scale with your tasks and real version upgrades.
goku update with goku CLI 0.7.0 or later (run goku update --cli first). After an update with an older CLI, or a new docker run, a function deployed from S3 gets its code back on its next invoke; a function uploaded as a zip file, and a layer, needs its code uploaded once more (aws lambda update-function-code, publish-layer-version). The invoke error tells you which./_goku/ (/_mimir/ keeps working), its DNS resolves *.localhost.goku.local and the configured hostname (add others with GOKU_DNS_EXTRA_SUFFIXES), and REST APIs are called at /restapis/…/_user_request_/ or /execute-api/….dig +tcp: Docker Desktop for Mac with kernel networking for UDP does not return UDP answers from containers.Update with Update now in the console, or run goku update --cli and then goku update. Running Goku 4.1 or later shows the update by itself.
4.2.3: RDS runs the database versions AWS offers today. Each engine version now runs exactly that version, and new databases default to PostgreSQL 18.6, MySQL 8.4.11 and MariaDB 12.3.3. Upgrade a database to a newer version from the console or with ModifyDBInstance: its data, users and passwords move with it, and if the new version does not come up the old one comes back on the untouched data. Databases created on earlier releases keep the engine that created their data and show the version they really run.
4.2.2: Every engine that cannot start now says why. An RDS instance, Redshift cluster, OpenSearch domain, EKS cluster, ElastiCache or MemoryDB cluster, DocumentDB or Neptune cluster, Amazon MQ broker or MWAA environment whose engine exits, runs out of memory or never becomes ready fails at once with the engine's own error, in its status, its events and the console, with a link to the engine's logs. Nothing stays creating forever, and PostgreSQL, MongoDB, RabbitMQ and Valkey keep their data in Docker volumes where the data folder does not keep owners and permissions (a Windows drive, for one). Thanks to the user who reported their RDS instances failing without a reason.
4.2.1: Amazon MSK now works on macOS and Windows: Kafka brokers keep their data in a Docker volume where the data folder is a Docker Desktop file share, and data from a cluster that already ran is copied over. Brokers are reachable from your machine at the localhost address GetBootstrapBrokers returns, a cluster whose broker cannot start moves to FAILED and says why, and goku services subscribe msk works. Thanks to the user who reported these. Update with goku update or Update now in the console.
Goku 4.2 answers 73% of the AWS operations it emulates, up from 42%: Glue, the core services and every data service are now complete. Workspaces let you run only the services you're working with today and pause or stop the rest to save memory. And the IAM Lab can show what any role in your real AWS account can do, using the AWS profile you already have.
Save the services you work with together, such as “Serverless day” (Lambda, ECS, DynamoDB) or “Data day” (Glue, S3, Athena), then Start, Pause, Stop or Switch to them as one. Pause keeps the containers' memory and uses no CPU; Stop frees the memory and keeps the data. How workspaces work.
IAM Lab → Role access connects to your AWS account with a profile you already have: access keys, SSO or assume-role, with no role to deploy. Pick a role to see every action it is allowed and the statement behind it, who can assume it in plain words, and its risks. Read-only, always. See a role's permissions.
goku ws starts, pauses, stops and switches workspaces from your terminal, and goku iam profiles, connect, roles and access ROLE show what a role can do. Nine new MCP tools let AI assistants do both, 34 in total.




Goku 4.2 adds 1,820 operations. Each one keeps real state and follows AWS's request and response shapes, errors and paging.

| Service | Goku 4.1 | Goku 4.2 |
|---|---|---|
| Glue | 107 / 300 | 299 / 300 |
| Systems Manager | 23 / 152 | 152 / 152 |
| CloudWatch Logs | 27 / 118 | 117 / 118 |
| IAM | 100 / 180 | 178 / 180 |
| Cognito user pools | 57 / 132 | 132 / 132 |
| CloudFormation | 27 / 90 | 89 / 90 |
| Athena | 13 / 70 | 70 / 70 |
| API Gateway (REST, HTTP and WebSocket) | 125 / 227 | 227 / 227 |
| Lambda | 54 / 88 | 88 / 88 |
| S3 | 71 / 116 | 116 / 116 |
| RDS | 39 / 164 | 164 / 164 |
| Redshift and Redshift Serverless | 87 / 210 | 210 / 210 |
| OpenSearch and the Elasticsearch Service API | 24 / 147 | 147 / 147 |
| EKS | 7 / 70 | 70 / 70 |
ECR, AWS Backup, ElastiCache, Transfer Family, Neptune, MSK, KMS, Secrets Manager, STS, CloudWatch, EventBridge, Step Functions, Kinesis, Firehose, ECS, DynamoDB, SNS, DocumentDB, MemoryDB and Lake Formation are complete too.
A few of the new operations:
Some answers changed to match AWS's behaviour, so code that relied on the old ones may notice:
GetCallerIdentity reports the caller: the IAM user of a user key, or the assumed-role or federated-user ARN of a session. Before, it always reported the account root.FilterLogEvents and metric filters use AWS's filter-pattern syntax instead of substring matching.DeleteUser refuses while the user still has credentials.creating until the engine accepts connections, EKS clusters created from 4.2 on check tokens against their access entries, and new MSK clusters default to TLS-only client traffic.Also in 4.2: images pushed to an ECR repository's URI with docker push now show up in DescribeImages and ListImages, with a registry front that keeps each account's and Region's images apart, and the last Mimir names users could see, such as the images and S3 buckets Goku creates, are Goku names now. Nothing needs to be rebuilt or re-created.
Goku 4.1 connected AI assistants to your local AWS. Goku serves the Model Context Protocol at http://localhost:8080/mcp, so Claude Code, Claude Desktop, Cursor, VS Code, Windsurf and any other MCP client can create and inspect resources, call any AWS API, read logs, take snapshots and inject faults, all against the sandbox on your machine. goku CLI 0.5.0 added goku mcp, a stdio bridge that starts Goku when it isn't running, and goku mcp config prints the setup for each client. Connect an AI assistant to Goku.
Mimir is now Goku, everywhere. Goku tells you when a new version is out and updates itself from the console, and everything it keeps, including your databases, lives in a folder on your disk. It also adds AWS Glue 6.0, DynamoDB vector search, AWS Lambda runtime parity, a rebuilt ECS, the IAM Policy Lab, a lightweight start with per-service subscriptions, a Resource Center, real engines behind a dozen more services, and console pages for 33 more services — 60 in total, on a backend of 73.
Every name is Goku now: the image tanujsoni027/goku, the goku container and command, GOKU_* settings, /_goku paths, goku.local host names and the goku-ca.pem file. The site moved to tanuj24.github.io/goku.
goku update, or Update now in the console, renames your mimir container to goku and moves it to the new image, and your data into the data folder.MIMIR_* setting, with the new name to use.The console shows when a new version is out. One click downloads it and replaces the running Goku with the same ports, settings and data; if the new version doesn't come up healthy within 3 minutes, Goku rolls back to the one you had. The check is one download of a static file, with nothing about you sent; GOKU_UPDATE_CHECK=off turns it off.
Buckets, tables, functions and the data of every engine (RDS PostgreSQL databases, OpenSearch, ECR, DocumentDB, MemoryDB, Amazon MQ and Redshift) live in ~/.goku/data (%LOCALAPPDATA%\Goku\data on Windows). On macOS and Windows, MSK and RDS MySQL keep their data in Docker volumes, which also survive restarts and updates, and so do PostgreSQL, MongoDB, RabbitMQ and Valkey where the data folder does not keep owners and permissions (a Windows drive, for one). Reset Docker, update or reinstall: start Goku with the same folder and resume where you left off. Existing data moves in by itself.
goku update moves your install to the data folder on the way; goku update --cli updates the CLI itself, verified against its published checksums; goku data shows, opens or migrates the folder. The CLI tells you, at most once a day, when a new Goku or CLI is out.
Install the CLI and run goku update: it saves your state, moves it into the data folder, renames your mimir container to goku and starts the newest Goku on the same ports. Your old mimir-data volume stays as a backup. Using plain Docker? See upgrading an existing install.
Glue 6.0 (Spark 4.1.1, Python 3.13, Java 17) and Glue 5.1, per job or notebook. AWS publishes no Glue 6.0 image, so Goku builds one locally the first time a 6.0 job runs — a one-time step of a few minutes. New jobs default to Glue 5.1, as on AWS.
The VectorIndexes and SearchVectors APIs with cosine, Euclidean and dot-product distance, and a console tab to search by text, by a similar item or by raw vector. Plus a rebuilt item explorer and a PartiQL editor.
Every runtime AWS Lambda supports today — Node.js, Python, Java, .NET, Ruby and provided.al2023 — on arm64 and x86_64, with AWS's lifecycle dates. Also provisioned concurrency, SnapStart, container-image functions and response streaming.
Self-healing services, rolling deployments and a circuit breaker with rollback; health checks, secrets and volumes; ECS Exec from the AWS CLI, Service Connect and AWS Cloud Map; and a browser terminal into running containers.
Answers “if I ship this IAM change, who gains or loses what access, and why?” before you ship it: effective access with evidence, policy drafts you can diff, simulate and size by blast radius, and IAM tests with regression reports — for Goku or, read-only, a real AWS account.
A new installation starts only the core services — S3, DynamoDB, Lambda, SQS, SNS, IAM/STS, CloudWatch, KMS, Secrets Manager and Systems Manager. Subscribe to more, or to a preset such as Containers, Databases or Analytics, on the Services page.
Everything Goku runs — databases, brokers, clusters, functions and build containers — with live memory and CPU. Pause, stop, restart or remove any of it from one place, with a memory budget, idle auto-stop and disk cleanup.
Bedrock through a local Ollama, Redshift and its Data API, Amazon MWAA on real Airflow, EMR Serverless on real Spark, MemoryDB on real Valkey, DocumentDB on real MongoDB and Amazon MQ on real RabbitMQ or ActiveMQ.
60 in total. New pages include Route 53, Elastic Load Balancing, CloudFront, CodeBuild, CodeDeploy, AWS Backup, AWS Config, Bedrock, Cost Explorer, Redshift, Amazon MWAA, CloudTrail, X-Ray and AWS Organizations.
New developer tools, in the console, the API and the goku CLI:
AccessDenied.TestState API, mock configurations and test suites with JUnit reports.goku CLI and a VS Code extension: services, readiness, snapshots, fault injection, IAM and Lambda debugging from your terminal or editor.Also new in 4.0: real engines behind a dozen more services — including Bedrock through a local Ollama, Redshift, Amazon MWAA on real Airflow, EMR Serverless on real Spark, MemoryDB, DocumentDB and Amazon MQ — and console pages for 33 more services, from Route 53 and CloudFront to Cost Explorer and AWS Organizations.
A real-time workflow graph: states color in as the execution progresses, Choice branches show their conditions, and Task nodes deep-link to the local resources they invoke.

A tour of the bundled web console: 27 of its 70 service consoles, grouped the way AWS groups them, plus Goku's own tools, with Workspaces and IAM Role access new in Goku 4.2. Click any one to see its page.

Storage · S3 console in Goku
The bundled web console covers 70 services, and every service page has a one-click “Load sample data” button, so you can explore a populated cloud in under a minute.

RDS starts a real PostgreSQL, MySQL or MariaDB container — connect with psql on localhost:5500. ElastiCache and MemoryDB run real Redis or Valkey, DocumentDB runs real MongoDB, Amazon MQ runs real RabbitMQ or ActiveMQ, OpenSearch is the real engine and EKS boots real k3s. Lambda functions run in real runtime containers.
Build REST, HTTP and WebSocket APIs in the console, attach a Lambda integration, and invoke them on localhost seconds later — no deploy pipeline, no waiting.
Real PySpark jobs and interactive notebooks with the real awsglue libraries, Glue 6.0 back to 2.0, with Hudi, Delta Lake and Iceberg. Step Functions Task states really invoke your local Lambdas.
Your buckets, tables, functions and jobs, and the data of every database and broker, live in a folder on your disk. Stop, update, reinstall or reset Docker: start Goku with the same folder and carry on.
Data endpoints live in a dedicated 5500–5689 range, so they never clash with a local Redis, Postgres or Elasticsearch. RDS, ElastiCache and Neptune use the published 5500–5524 ports; the rest bind on your host automatically.
CloudFormation stacks actually provision the buckets, queues and topics in your template, and every container-backed service streams logs to its AWS-convention CloudWatch log group, exactly like the real cloud.
Zero code changes: point your AWS CLI, SDK or app at http://localhost:4566, region us-east-1, with any local credentials such as test / test — or let goku env set it all for you.
Step-by-step walkthroughs with commands you can paste, for the things developers most often run locally.
Install Goku, point the AWS CLI at localhost:4566 and create S3 buckets, DynamoDB tables, SQS queues and Lambda functions on your machine.
Create, invoke and debug functions in real runtime containers, read their CloudWatch logs and stop on breakpoints in VS Code or JetBrains IDEs.
Tables, items, GSI queries, PartiQL, Streams and vector search on localhost, with your data kept between restarts.
Glue 6.0 and 5.1 jobs and notebooks on real Spark with the awsglue libraries, a local Data Catalog and S3.
Give Claude Code, Claude Desktop, Cursor, VS Code or Windsurf 34 tools and any AWS API call on a local sandbox, never your real AWS account.
Save the services you work with as workspaces and start, pause, stop or switch them together, so only what you use today takes memory.
Connect the IAM Lab with an AWS profile you already have and see a role's effective permissions, who can assume it and its risks, read-only.
Only the name. Goku is the same product, renamed in Goku 4.0, where every name became Goku: the image tanujsoni027/goku, the goku container and command, GOKU_* settings, /_goku paths and goku.local hostnames. Existing installs move over automatically, and the old Mimir names keep working for now, so start using the goku names. See old and new names.
Install Goku and point your AWS CLI or SDK at http://localhost:4566, region us-east-1, with any local credentials such as test / test — goku env sets these for you. No account and no sign-up. Once the image and any runtime images your services need are pulled, everything runs on your machine without internet access. Walkthrough: how to run AWS locally.
Yes. Functions run in real runtime containers for every runtime AWS Lambda supports today, with logs, versions, aliases, function URLs and event source mappings. Step through a handler from VS Code, IntelliJ, PyCharm or GoLand. See testing AWS Lambda locally.
Yes. Save the services you work with together as a workspace, for example Lambda, ECS and DynamoDB, and start, pause, stop or switch them as one from the console, goku ws or MCP. Pause freezes the containers: they keep their memory and use no CPU. Stop frees the memory and keeps the data. See running only the services you need.
Yes. The IAM Lab connects to your AWS account with an AWS profile you already have (access keys, SSO or assume-role), scans its roles with read-only calls and shows for any role every action it is allowed, who can assume it and its risks. Nothing in the account changes, and credentials are held in memory only. See what an AWS IAM role can do.
Yes. Everything Goku keeps, databases included, lives in a data folder on your disk (~/.goku/data, or %LOCALAPPDATA%\Goku\data on Windows). It survives stops, updates, reinstalls and Docker resets. Environment snapshots can save and restore the whole local cloud.
Yes. The goku CLI installs natively on macOS, Linux and Windows and runs Goku in Docker. The image is multi-arch (arm64 + amd64), so it runs natively on Apple Silicon.
Docker: Docker Desktop on macOS and Windows, and Docker Engine or Docker Desktop on Linux. See all install options.
No. Goku is a local development and testing sandbox — build and iterate locally, then deploy to real AWS for production.