Service mapping
This page maps every service on the Crusoe Agent Platform to its closest equivalent on AWS, GCP, and Azure. Use it to translate what you already know — then jump to the per-cloud guides for the full vocabulary: AWS, GCP, Azure.
We are in alpha, and we say so
The big clouds have things we do not: many regions, SLAs, compliance programs, and giant marketplaces. This page is honest about that. What we offer instead is one small platform where every service works the same way, every docs page is copy-paste runnable, and nothing is split across competing product generations.
The master table
Each row is one job you might want done. The first column links to our docs for that service.
| Crusoe Agent Platform | AWS | GCP | Azure | Honest notes |
|---|---|---|---|---|
| Agent Engine — deploy agent code as a serverless chat endpoint | Amazon Bedrock AgentCore (Bedrock Agents "Classic" is closed to new customers) | Vertex AI Agent Engine (now part of the Gemini Enterprise Agent Platform) | Microsoft Foundry Agent Service | We run your unmodified ADK, LangGraph, or CrewAI code — one deploy path, no "Classic vs v2" fork. The clouds have more model choice and a no-code option (Azure) that we do not. |
| Serverless — container in, scale-to-zero HTTPS endpoint out | AWS App Runner (closed to new customers) or AWS Fargate | Cloud Run | Azure Container Apps | Cloud Run is the closest match, down to the word "revision". The clouds offer far more regions and instance sizes. |
Functions — one handle(event) file becomes an HTTPS endpoint | AWS Lambda | Cloud Run functions | Azure Functions | Four runtimes here (Python, Node.js, Go, Ruby). No API Gateway to wire up — every function gets a URL. The clouds have much larger trigger and binding catalogs. |
| VectorDB — collections, points, similarity search | Amazon OpenSearch Service vector engine (k-NN) | Vertex AI Vector Search | Azure AI Search vector store | Ours is a dedicated vector database. AWS and Azure bolt vectors onto a search engine; GCP's is purpose-built but recently changed its whole data model (1.0 to 2.0). |
| MemoryStore — in-memory key-value store | Amazon ElastiCache or Amazon MemoryDB | Memorystore for Redis and Valkey | Azure Managed Redis (Azure Cache for Redis is retiring) | Speaks the Redis/Valkey protocol, so your existing client libraries work unchanged. One product — not a cache-vs-database split (AWS) or a retiring-and-successor pair (Azure). |
| Pub/Sub — topics and subscriptions | Amazon SNS + SQS + EventBridge | Cloud Pub/Sub (plus Eventarc) | Azure Service Bus + Event Grid | One messaging service. AWS and Azure each split this job across two or three products and publish official "which one do I use" decision guides. We never make you choose. |
| Durable Execution — workflows that survive restarts | AWS Step Functions | GCP Workflows | Azure Durable Functions | You write workflows as ordinary code, not a JSON or YAML state language. Step Functions has richer visual tooling and hundreds of built-in AWS service integrations that we do not. |
| Code Sandbox — run untrusted code in isolation | Bedrock AgentCore Code Interpreter | Gemini code execution tool | Azure Container Apps dynamic sessions | Every execution here gets a single-use pod that is destroyed afterward — no session pools, identifiers, or lifecycles to manage. The clouds support longer-lived, stateful sessions, which we deliberately do not. |
| Secrets Manager — store credentials, inject them into workloads | AWS Secrets Manager | GCP Secret Manager | Azure Key Vault | Secrets here are write-only values injected as environment variables. No vault objects or version stages to learn first — but also no automatic rotation and no version history. |
| Sessions — per-conversation history for agents | Bedrock AgentCore Memory (short-term) | Vertex AI Sessions | Foundry threads | Built into every deployed agent; the platform replays the history to the model each turn. Not a separate product to enable. |
| Memory — long-term memory agents can search | Bedrock AgentCore Memory (long-term), Bedrock Knowledge Bases | Vertex AI Memory Bank | No direct equivalent (ground with Azure AI Search) | You decide when a conversation becomes long-term memory ("memorize"); agents search it with a built-in tool. Anything memorized is readable by every later caller of that agent — we say that plainly. |
| MCP Servers — host Python tools as a Model Context Protocol endpoint | Bedrock AgentCore Gateway (wraps APIs as MCP tools) | No direct managed equivalent | No direct managed equivalent | Every publish creates an immutable version; rollback is one command. |
| Projects and access — who can touch what | AWS IAM + AWS Organizations | GCP projects + IAM | Azure subscriptions and resource groups + Entra ID RBAC | Two roles (admin, member) per project — far simpler than an IAM policy language, and far less expressive. Platform staff cannot read your data without a visible, expiring, audited break-glass grant. |
| Service accounts and API keys — machine identities | IAM roles + access keys | GCP service accounts + keys | Entra service principals and managed identities | Modeled on GCP: a service account belongs to one project and holds a role like a person would. Key secrets are shown exactly once. |
| Quotas and audit — capacity limits and history | Service Quotas + AWS CloudTrail | GCP quotas + Cloud Audit Logs | Azure quotas + Activity Log | Every project gets a live quota view and an always-on audit log readable by every member — including any break-glass access taken by the provider. No export or retention controls yet. |
| Embed chat — put your agent on a website | No direct equivalent | No direct equivalent | No direct equivalent (Foundry publishes to Copilot instead) | One rotatable public key plus an origin allowlist gives any website a chat widget for your agent. |
What the big clouds have that we don't
Honesty first. If any of these are hard requirements for you today, a big cloud is the right choice — and we would rather tell you that here than have you find out later.
- Regions and zones. AWS, GCP, and Azure run dozens of regions with availability zones. The platform runs in one place today. There is no region picker.
- SLAs and compliance. No uptime SLA, and no published compliance programs (nothing like VPC Service Controls, CMEK, or HIPAA attestations).
- Marketplace and catalog breadth. The clouds have hundreds of adjacent services and marketplace listings. We have the services in the table above, and that is all.
- Model catalogs. Bedrock, Vertex, and Foundry expose large first-party model catalogs. Agents here talk to an OpenAI-compatible Crusoe Managed Inference endpoint configured by your administrator.
- Identity maturity. Sign-in is email and password. There is no MFA, no SSO or federation, and sign-in tokens last 12 hours and cannot be revoked early. Accounts are created by an admin or an invitation link — there is no self-service sign-up.
- Secret lifecycle features. No automatic rotation, no version history, no soft-delete. You rotate a secret by writing a new value, which rolls a new revision of the workload.
- Guardrails and content safety. Bedrock Guardrails and Azure's responsible-AI layer have no equivalent here yet.
- Historical monitoring. Our metrics are live counts only — there are no historical charts. Log history is kept 14 days by default. The audit log has no export, retention controls, or date-range filter yet.
- GPU classes. The clouds let you pick specific GPU instance types. The Agent Platform has no GPU instance-class selection today.
- A public API hostname. There is no public API endpoint yet. The web console at https://console.codyhill.dev is the front door; automation reaches the API through the
$CAI_APIaddress your administrator provides (often a port-forward). See API authentication.
Where this platform is genuinely better
These are not marketing claims — each one contrasts with a documented, current weakness of a competitor product.
- One product per job, one generation of each product. AWS runs Bedrock Agents "Classic" (closed to new customers, docs still live) alongside AgentCore. App Runner is closed to new customers. Azure Cache for Redis is retiring in favor of Azure Managed Redis. GCP has renamed its agent runtime twice, and its API paths still use the old name. Here there is exactly one way to deploy an agent, and we promise: if we ever rename anything, the old name keeps working and the docs will say so.
- No wiring between services. A function gets an HTTPS URL without an API Gateway. An agent tool is a Python function, not an OpenAPI schema plus a Lambda. Sessions and memory come with every agent instead of being separately enabled sub-products.
- Simple, honest security defaults. Asking about a project you have no grant on returns 404, never 403, so nothing about your projects can be discovered by guessing. Provider staff hold no standing access to your data — the only path in is a break-glass grant that appears in your member list and your audit log.
- No billing surprises. There is no usage-based metering today. Capacity is governed by per-project quotas you can see live. Compare that to a single Vertex agent turn that can produce four billed SKUs, or AgentCore's roughly twelve separately metered components.
- Runnable docs. Every command in these docs shows its expected output, and the quickstarts state the time they take. That is a deliberate answer to quickstarts elsewhere that require account setup, IAM policy authoring, or a second VM before "hello world".
Next steps
- Coming from AWS? Read the AWS guide.
- Coming from GCP? Read the GCP guide.
- Coming from Azure? Read the Azure guide.
- New to all of it? Start with core concepts and the agent quickstart.