swarmd vs AWS · Azure · Google Cloud
More control than the hyperscalers. On your prem, or on theirs.
AWS, Azure and Google Cloud each govern the agents that live on their cloud, across a handful of separate services. swarmd governs every agent, MCP server and model gateway you run — in one control plane, on your own infrastructure or any of theirs — and lets you move between them whenever you like.
swarmd80of 80 built in
Every capability, on your prem or ours
AWS17of 80 built in
39 partial · 24 not offered
Microsoft Azure22of 80 built in
39 partial · 19 not offered
Google Cloud17of 80 built in
43 partial · 20 not offered
Counted from the table below. Reviewed against each provider’s public documentation for generally available agent services.
More of the controls you need
Human review queues, hash-chained audit, persona dashboards, agent-aware detections and EU AI Act mapping — built in, not assembled from five services.
Deploy on your prem
The whole platform from one Helm chart, in your data centre or an isolated network. Your servers, your keys, your jurisdiction.
No lock-in, in any direction
Run on AWS, Azure or Google Cloud today and move off tomorrow — or the reverse. Open protocols, portable policies, your data.
Capability by capability
The whole platform, against the big three.
Every row is a capability swarmd ships today, grouped the way the features catalogue is. Where a hyperscaler offers it, we say so — and name the service you would use.
Built in — Offered as a product feature.Partial — Narrower, in preview, or assembled from several separate services.Not offered — Not a product feature — you would build it.
Showing 80 of 80 capabilities.
| Capability | swarmd | AWS | Microsoft Azure | Google Cloud |
|---|---|---|---|---|
| Deployment·Where it runs | ||||
Runs entirely in your own data centre The full platform, not a cut-down edge edition. | Built inOne Helm chart on your Kubernetes | Not offeredAgentCore runs in AWS regions only | PartialFoundry Local on Azure Local; Foundry Agent Service is cloud-only | PartialGemini on Google Distributed Cloud; the agent platform is cloud-hosted |
Isolated and air-gapped networks | Built inMirror the images; offline licence validation by arrangement | Not offered | PartialAzure Local disconnected operations, separate product | PartialGoogle Distributed Cloud air-gapped, separate product |
Runs on any cloud — including AWS, Azure and Google Cloud | Built inEKS, AKS, GKE or any conformant Kubernetes | Not offeredAWS only | Not offeredAzure only | Not offeredGoogle Cloud only |
Same product managed or self-hosted Same features, same SDKs, same API — only the operator changes. | Built inswarmd Cloud or your cluster | Not offered | PartialOn-prem editions carry a narrower feature set | PartialOn-prem editions carry a narrower feature set |
Data stays in the jurisdiction you choose | Built inAny site you run, plus your own databases | PartialWithin the AWS region list | PartialWithin the Azure region list | PartialWithin the Google Cloud region list |
You decide when to upgrade | Built inUpdates on your terms, never forced | PartialAPIs are versioned; the service updates on AWS’s schedule | PartialAPIs are versioned; the service updates on Microsoft’s schedule | PartialAPIs are versioned; the service updates on Google’s schedule |
Source code held in independent escrow | Built inReleased to you if we can’t support it | Not offered | Not offered | Not offered |
A managed option if you’d rather not run it | Built inswarmd Cloud, in your region | Built in | Built in | Built in |
| No lock-in·Portability | ||||
Governs agents wherever they run On-prem, on any cloud, or a mix — registered by URL. | Built inAny agent that speaks A2A, or via our SDKs | PartialAny framework, but hosted on or routed through AWS | PartialAgent 365 sync with AWS and Google Cloud in preview | PartialExternal agents via A2A in the agent registry |
Any LLM provider behind one gateway | Built inOpenAI, Anthropic, Mistral and more | PartialBedrock catalogue; others via your own integration | PartialFoundry catalogue, or bring your own AI gateway | PartialModel Garden catalogue |
Bring any agent framework | Built inGoogle ADK and LangChain SDKs, or plain A2A | Built in | Built in | Built in |
Open protocols: A2A and MCP | Built in | Built in | Built in | Built in |
Policies in an open, portable language | Built inRego, round-tripped with the visual builder | PartialCedar, with an AgentCore-specific schema | Not offeredAPI Management XML and Azure Policy | Not offeredModel Armor templates |
Audit trail you can verify outside the platform | Built inExport with a hash-chain verification report | PartialCloudTrail digests cover API calls, not agent messages | PartialLog export without an integrity chain | PartialLog sinks without an integrity chain |
Leave without rewriting your agents | Built inAgents keep their code; point them elsewhere | PartialAgents built on AgentCore services need porting | PartialAgents built on Foundry services need porting | PartialAgents built on Agent Engine services need porting |
Built on widely used open-source infrastructure | Built inKubernetes, Postgres, Keycloak, ClickHouse | Not offeredProprietary managed services | Not offeredProprietary managed services | Not offeredProprietary managed services |
| Control plane·One portal for the whole estate | ||||
One portal for agents, MCP servers, LLM gateways and channels | Built in | PartialSplit across AgentCore, Bedrock, CloudWatch and IAM | PartialSplit across Foundry, Agent 365, API Management and Purview | PartialSplit across Gemini Enterprise, Agent Platform, Apigee and SCC |
Closed by default — every link an explicit grant | Built inAgents can’t self-discover or self-connect | PartialThrough IAM and gateway targets | PartialThrough Entra Agent ID and RBAC | PartialThrough IAM |
Configure once, agents pick it up without redeploys Subscriptions, credentials, routes and freezes. | Built in | Partial | Partial | Partial |
Live network map of every agent, tool, gateway and user | Built inHealth-sorted topology | PartialPer-trace service maps | PartialAgent 365 relationship views | PartialPer-trace views |
Test the policy on any connection, from the map | Built in | Not offered | Not offered | Not offered |
Agent marketplace | Built inVerified agents, one-click connect | Built inAWS Marketplace | Built inAgent Store | Built inAgent marketplace |
Curated and community MCP catalogue | Built inWith uptime and latency metrics | PartialGateway targets; no community tier | PartialTool catalogue; no community tier | PartialAPI registry; no community tier |
| Agents·Lifecycle and orchestration | ||||
Agent registry and inventory | Built in | Built in | Built inAgent 365 registry | Built inAgent registry |
Killswitch for a single agent, with a reason | Built in | PartialStop the runtime or revoke IAM | Built inBlock in Agent 365 | PartialDisable via IAM or the registry |
Tenant-wide lockdown during an incident Freeze every agent, caller and webhook at once. | Built in | Not offeredAssemble from IAM and org policies | Not offeredAssemble from Entra and Azure Policy | Not offeredAssemble from IAM and org policies |
Zero-downtime credential rotation | Built inWithout changing the client ID | Built inSecrets Manager | Built inKey Vault | Built inSecret Manager |
Skills inventory you can govern | Built in | PartialGateway tool listing | Partial | Built inSkills and tools registry |
Live task board of every agent task | Built inKanban by lifecycle state, with history | Not offered | Not offered | Not offered |
Per-agent uptime, latency and error metrics | Built in | Built inAgentCore Observability | Built inFoundry observability | Built inCloud Monitoring |
Lifecycle audit trail per agent | Built inEvery state change, in order | PartialControl-plane calls in CloudTrail | PartialControl-plane calls in the Activity Log | PartialControl-plane calls in Cloud Audit Logs |
Freeze, version and rotate MCP servers | Built in | Partial | Partial | Partial |
| Connectivity·Gateways, models and channels | ||||
LLM gateway with multi-provider failover | Built in | PartialCross-region inference within Bedrock | Built inAPI Management AI gateway | PartialApigee, configured by you |
Route models by agent metadata tag | Built inPriority-ordered, with fallbacks | Not offered | PartialHand-written API Management policies | PartialHand-written Apigee policies |
Token usage and cost per agent | Built inGateway and policy tokens, priced | PartialCost allocation by inference-profile tags | Built inAPI Management token metrics | PartialBilling export by label |
MCP auth: bearer, API key, OAuth2 and user-delegated | Built in | Built inAgentCore Identity | Built in | Partial |
Microsoft Teams and OAuth-backed channels | Built in | Partial | Built in | Partial |
| The gatekeeper·Policy and enforcement | ||||
One policy model across agent, MCP and LLM traffic | Built in | PartialSeparate engines: Policy, Guardrails, IAM | PartialSeparate engines: API Management, Content Safety, Purview | PartialSeparate engines: Model Armor, Apigee, IAM |
Visual, no-code policy builder | Built inDrag and drop | PartialNatural-language authoring to Cedar | PartialPurview DLP rules; gateway policies are XML | PartialModel Armor templates |
Policy as code, round-tripped with the visual builder | Built inRego to visual and back | PartialCedar, one direction | Not offered | Not offered |
Block, mask, warn, log or hold for human review Five actions on every exchange. | Built in | PartialAllow, deny, block or mask | PartialAllow, block or annotate | PartialInspect or block |
PII detection and redaction | Built inPresidio (23+ types) and AWS Comprehend (21) | Built inBedrock Guardrails | Built inAzure AI Language and Purview | Built inSensitive Data Protection |
Regex, rate limits and tool restrictions | Built in | Built in | Built in | Partial |
Target request and response legs separately | Built in | Partial | Partial | PartialFirst prompt and final response |
11 cascade scopes with an inherited policy chain Tenant, agent, user, channel, MCP, gateway and more. | Built in | PartialAccount and organisation levels | PartialManagement-group and resource levels | PartialOrganisation, folder and project levels |
Versioned policy groups with owners and validity windows | Built inHash-tracked versions, full history | PartialVersions only | PartialVersions only | PartialVersions only |
Validate and preview before enforcing | Built in | Partial | Partial | Partial |
Log of every request a policy intercepted | Built in | Partial | Partial | Partial |
| Human-in-the-loop·People stay in control | ||||
Central approvals queue for flagged exchanges | Built in | Not offeredBuild it with Step Functions | Not offeredBuild it with Logic Apps | Not offeredBuild it with Workflows |
Review with the full context and linked trace | Built inUser, agent, policy, matched content | Not offered | Not offered | Not offered |
Every approve or reject recorded with who, when and why | Built in | Not offered | Not offered | Not offered |
Review SLAs, outcomes and escalations | Built in | Not offered | Not offered | Not offered |
Reviewer leaderboard and override audit | Built in | Not offered | Not offered | Not offered |
| Observability·See everything, interfere with nothing | ||||
Every exchange logged with rich context | Built in | Built inInvocation logging and CloudWatch | Built inAzure Monitor | Built inCloud Logging |
Distributed tracing with OpenTelemetry | Built inOTLP ingest from agent SDKs | Built in | Built in | Built in |
P1–P5 alerts with acknowledge, resolve and reopen | Built in | PartialAlarms; incident lifecycle is a separate service | PartialAlerts; incident lifecycle is a separate service | PartialAlerts; incident lifecycle is a separate service |
Mute and downtime windows, with who set them | Built in | Partial | Built inAlert processing rules | Built inSnoozes |
Search, stats and CSV export across logs, traces and alerts | Built in | PartialSeparate consoles | PartialSeparate consoles | PartialSeparate consoles |
| Detection·Agent-aware anomaly detection | ||||
Dormant agents or users suddenly reactivating | Built in | PartialGuardDuty, for IAM principals | PartialDefender for Cloud AI threat protection | PartialSecurity Command Center AI Protection |
An agent calling a tool it has never used before | Built in | Not offered | Partial | Partial |
An agent binding to a new MCP server | Built in | Not offered | Not offered | Not offered |
Sensitive data flagged in live traffic | Built in | Built in | Built in | Built in |
Coverage gaps: agents with no policy, shadow policies, unmonitored agents | Built in | Not offered | PartialAgent 365 posture | PartialAI security posture |
| Compliance·Tamper-evident audit and the EU AI Act | ||||
Hash-chained, signed log of agent exchanges | Built in | PartialCloudTrail integrity covers API calls only | PartialImmutable storage, not chained | PartialLocked log buckets, not chained |
RFC 3161 third-party timestamps | Built in | Not offered | Not offered | Not offered |
One-click integrity check over any time range | Built in | PartialCLI digest validation | Not offered | Not offered |
Compliance export with a verification report | Built in | Partial | Partial | Partial |
EU AI Act obligations per system, by tier, role and use case Checked against live configuration, not a questionnaire. | PartialAudit Manager frameworks | PartialPurview Compliance Manager templates | PartialCompliance Manager frameworks | |
Article 50 disclosure added to agent replies in-line | Built in | Not offered | Not offered | Not offered |
Article 73 incident clocks raised from monitoring alerts | Built in | Not offered | Not offered | Not offered |
| Insights·Dashboards for every stakeholder | ||||
Seven persona dashboards, CIO to domain owner | Built in | Not offeredBuild in QuickSight or CloudWatch | PartialAgent 365 and Purview reports | Not offeredBuild in Looker |
Control effectiveness: share blocked, masked, warned or held | Built in | PartialGuardrail metrics | PartialContent Safety metrics | PartialModel Armor metrics |
Leaderboards for agents, users, MCP servers and reviewers | Built in | Not offered | Not offered | Not offered |
Adoption analytics: WAU/MAU, time to first value | Built in | Not offered | PartialAgent 365 usage | PartialGemini Enterprise analytics |
Error-budget burn and p95/p99 SLOs per agent | Built in | PartialGeneric SLOs, not per agent | PartialGeneric SLOs, not per agent | PartialGeneric SLOs, not per agent |
| Access & admin·Identity and administration | ||||
SSO with Entra ID, Google Workspace and Okta | Built in | Built in | Built in | Built in |
IdP group mapping and auto-provisioning | Built in | Built in | Built in | Built in |
A distinct identity for every agent | Built inOAuth client per agent | Built inAgentCore Identity | Built inEntra Agent ID | Built in |
Groups and fine-grained RBAC | Built in | Built in | Built in | Built in |
Chat with any agent or gateway to test it | Built inBuilt-in conversation UI | PartialAgents hosted on AWS | PartialAgents hosted on Azure | PartialAgents hosted on Google Cloud |
No lock-in
Move to them, off them, or between them.
A cloud’s own agent services are built to keep your agents on that cloud. swarmd is built to sit above whichever infrastructure you choose this year — and the next.
On-premCloud
Moving onto a hyperscaler
Run swarmd on EKS, AKS or GKE with the same Helm chart. Your policies, grants, audit trail and dashboards come with you unchanged.
CloudOn-prem
Moving off one
Bring the platform back into your data centre. Agents keep their code and protocols; only the URL they talk to changes.
Cloud ACloud B
Running across several
One control plane for agents on AWS, Azure, Google Cloud and your own servers — one policy set, one audit trail, one approvals queue.
What comes with you
- Policies as Rego
- Audit trail with its hash chain and verification report
- Agents on open A2A and MCP protocols
- Traces on OpenTelemetry
- Data in your own Postgres and ClickHouse
- Identity in your own IdP
Better together
Where the hyperscalers lead.
swarmd is a control plane, not a cloud. For the things below, the hyperscalers are the right choice — and swarmd governs the agents that use them, from wherever you run it.
- Model hosting and trainingFoundation-model catalogues, fine-tuning and managed inference at scale. swarmd routes to their models; it doesn’t host them.
- Compute and acceleratorsGPUs, TPUs and Trainium. Run swarmd next to them on EKS, AKS or GKE if that’s where your agents live.
- Content-safety classifiersPrompt-shield and harmful-content models. Keep using them; swarmd puts its own policy, review and audit around every exchange.
How we compared: the cloud columns reflect each provider’s publicly documented, generally available agent services as of — chiefly Amazon Bedrock AgentCore and Guardrails; Microsoft Foundry, Agent 365, API Management and Purview; and Google Gemini Enterprise Agent Platform, Model Armor and Apigee. “Partial” includes capabilities in preview or that need several services combined. These services change often; check the providers’ documentation for the latest. Spotted something out of date? Tell us.
AWS, Amazon Bedrock, Microsoft, Azure, Google Cloud and related names are trademarks of their respective owners. swarmd is not affiliated with or endorsed by them.
Run it beside them, or instead of them
Book a demo, or start a free 3-month trial on your own infrastructure — on-prem, or on AWS, Azure or Google Cloud.
Control the exchange
Integrate once. Govern every agent you route through swarmd. Keep the evidence from day one.
Scale with your team
Pick a plan built for growing teams and regulated environments