Real Docker-backed engines where fidelity matters, fast in-process emulation everywhere else — including DynamoDB SearchVectors and an AI suite in every cloud. Point SDKs, CLIs, Terraform, and CDK at localhost:4566, :4577, and :4588 — SigV4 and fake credentials welcome.
Endpoint localhost:4566 · same SDKs, CLIs, and IaC you use in production.
Endpoint localhost:4577 · Azure CLI, SDKs, and Resource Manager templates.
Endpoint localhost:4588 · gcloud CLI, client libraries, and Terraform providers.
= real Docker-backed engine · everything else runs in-process
Built-in control planes that turn all three clouds into a lab — drive them from an AI agent over MCP at /_loopback/mcp, the unified console, or direct /_loopback/* curl. Each lab is inert until you switch it on; together they cover cost projection, chaos injection, virtual time, state snapshots, IAM why-403 diagnostics, multicloud interconnect authorize gates, and an offline GenAI gateway across Bedrock, Azure OpenAI, and Vertex.
A built-in Model Context Protocol server lets AI agents build real infra, inspect every cloud, and drive every lab — zero spend, zero blast radius. 42 AWS / 24 Azure / 19 GCP tools: inspect, provision, chaos, clock, cost, snapshots, interconnect, GenAI.
Connect with claude mcp add --transport http loopback http://localhost:4566/_loopback/mcp. Tools return structured JSON mapped to /_loopback/* and cloud-native APIs — agents chain create-bucket, set-chaos, advance-clock, and export-snapshot without parsing CLI output.
Bedrock Converse / ConverseStream / InvokeModel / InvokeModelWithResponseStream on :4566, Azure OpenAI chat completions (including SSE streaming) on :4577, and Vertex predict / streamGenerateContent on :4588 — one offline gateway with deterministic canned responses, Ollama backends, or token-cost simulation wired to the Cost Lab.
Grok 4.6 on Bedrock (xai.grok-4.6) and Vertex (grok-4.6 / publishers/xai). Azure has no official Grok — a real coverage gap agents must handle. Zero API keys, zero egress.
Inject latency and native-shaped errors into any service via /_loopback/chaos — ThrottlingException on DynamoDB, 503s on Lambda, timeouts on SQS — in any of the three clouds.
Chaos also applies to multicloud interconnect links: sever or slow an authorized attachment to test retries, circuit breakers, and failover without real network partitions. Clear faults with one DELETE.
A virtual clock per cloud via /_loopback/clock/advance and freeze. Fast-forward and STS tokens expire, DynamoDB TTL items sweep, EventBridge schedules fire, SQS visibility timeouts elapse, and Azure hidden queue messages reappear.
Freeze pauses all time-dependent logic for deterministic assertions. The same clock drives Cost Lab accrual — advance thirty days and billing math updates in the same step as expiry events.
A live per-service cost meter at /_loopback/costs and what-if projector at /_loopback/costs/whatif — S3 operations, Lambda invocations, DynamoDB RCUs, GenAI tokens, and Azure/GCP analogs priced against published tables.
Wired to the virtual clock: watch a month of spend accrue in seconds, or project scaled spend over six months before an agent provisions another bucket.
Export the entire emulator's state via /_loopback/state/export and restore with import — git for cloud state across all three providers. Break everything in a test, then rewind.
/_loopback/why-403 replays SigV4 denials against identity policies, resource policies, and permission boundaries — returning the exact statement index and failing condition key, not a generic AccessDenied.
Model cross-cloud connectivity like AWS Interconnect — hub on localhost:4566, Azure and GCP gateways on :4577 and :4588. Cross-cloud calls return 403 until authorized via POST /_loopback/interconnects/authorize.
Then S3 talks to Blob Storage, Pub/Sub reaches SQS, and chaos can drop the link mid-test — explicit trust boundaries without VPC peering, ExpressRoute, or Cloud Interconnect setup.
AWS, Azure, and GCP on your machine in seconds — no account, no auth token, no spend.
Loopback Cloud