Short answer. Amazon Bedrock AgentCore is a set of managed AWS services that developers use to build, deploy and operate their own agents. xpander is an AI agent control plane. It runs Claude Code, Codex, OpenCode and agents you describe in plain language, on the models your admins allow. One set of access, approval and budget rules covers all of them. It runs in xpander's cloud, in your own Kubernetes cluster, or air-gapped. Choose AgentCore when an AWS team wants managed parts for custom agents. Choose xpander when the company needs one place to run and govern its agents, inside or outside AWS.
What each product is
AWS describes AgentCore as "an agentic platform for building, deploying, and operating highly effective agents securely at scale using any framework and foundation model." It has 13 services, including Harness, Runtime, Gateway, Identity, Memory, Policy, Observability, Evaluations and Registry, that a team combines into its own agent stack.
xpander is one platform. Its parts share the same sign-in, rules and records. The agent runtime runs the agents. The AI Gateway carries every model call. The MCP Gateway gives each person's AI tools their approved tools. Employees use agents on the web, in the desktop app, in Slack and in Teams.
Where AgentCore is strong
Open on framework and model. AgentCore runs agents built with CrewAI, LangGraph, LlamaIndex, Google ADK, the OpenAI Agents SDK, Strands Agents or your own code, on "any foundation model in or outside of Amazon Bedrock".
Ready parts for builders. A managed harness, a separate microVM for each session, Browser, Code Interpreter and Evaluations.
Policy outside the agent's code. AgentCore Policy checks every tool call that passes through AgentCore Gateway against rules written in Cedar, and logs each decision.
Deep AWS fit. VPC connectivity across its services, AWS GovCloud (US-West) availability, and pay-per-use pricing with no upfront commitment.
Side by side
xpander | Amazon Bedrock AgentCore | |
|---|---|---|
What it is | A ready AI agent control plane: agent runtime, AI gateway, MCP gateway and an app for employees | Managed services for teams that build their own agents |
Where it runs | xpander cloud, your own Kubernetes cluster in any cloud or on-prem, or air-gapped | AWS Regions, including AWS GovCloud (US-West) |
Self-hosted or air-gapped | Yes. Air-gapped installs run every component in your network, with no connection to xpander cloud | Not documented |
Models | The models your admins allow, through one AI gateway, including open-weight models on your own GPUs | Any model in or outside Amazon Bedrock |
Coding agents | Claude Code, Codex and OpenCode run as governed agents, picked per conversation | AWS has published how to host coding agents such as Claude Code yourself on AgentCore Runtime |
Tool policy | Approved tools only, per-person sign-in, and approvals that wait for a named person | Cedar policies on calls through AgentCore Gateway |
Spend control | Budgets in credits that warn or block per organization, group, member, API key, agent, turn or end user | Per-caller limits on requests, tokens (inference targets only) and concurrent requests at the Gateway, session limits in Policy, and a token limit per harness call. No money budget per person or team is documented |
Record | An audit log of sign-ins, setting changes and every tool call, plus a log of every model call. The audit log exports to your OpenTelemetry collector or storage bucket | Logs and traces in Amazon CloudWatch, in OpenTelemetry format, which you can send to tools such as Datadog, Dynatrace and Langfuse |
Who uses it day to day | Developers, IT and every employee, on the web, in the desktop app, in Slack and in Teams | Developers who build and operate agents |
Pricing | xpander cloud: credits for the work agents do, with unlimited seats. Self-hosted: an annual license, from 50 agents, with your own model keys | Pay per use, metered per service |
The key differences
Parts or a finished platform. AgentCore Harness defines an agent from configuration, and Policy can require a recorded human approval. AgentCore has no app for employees, and it documents no money budget per person or team. The team builds those parts and joins the services into one platform. xpander ships them joined, so a company can put Claude Code, Codex and its own agents under one set of rules without building the platform first.
Where the rules apply. AgentCore Policy "intercepts all agent traffic through Amazon Bedrock AgentCore Gateways". Calls that do not go through that gateway are outside it. In xpander, every model call goes through the AI Gateway. Agents reach only the tools you approve, with per-person sign-in and approvals, and the audit log records each shell or file action an agent runs. So one set of access, approval and budget rules covers every agent that runs on xpander.
Inside AWS or anywhere. AgentCore runs in AWS Regions. xpander runs in its own cloud, in your Kubernetes cluster, or air-gapped with models in your network. xpander's reference self-hosted installs run on Amazon EKS, and the xpander AI Gateway can call Amazon Bedrock models.
How a coding agent runs under xpander
An admin adds providers and models in Platform > AI governance > Models.
A person picks Claude Code, Codex or OpenCode and a model in the Harness and model picker.
The agent calls the AI Gateway with an xpander key or session token, never a provider key.
A call over a blocking budget is refused and says so. The call is in Logs, and the action is in the audit log.
How to choose
If your situation is... | Choose |
|---|---|
Your agents are built with LangGraph, CrewAI, Strands or your own code | AgentCore |
An AWS team wants managed parts to build custom agents | AgentCore |
You want Cedar policies on tool calls inside AWS | AgentCore |
You need one place to run and govern Claude Code, Codex and custom agents for the whole company | xpander |
AI must run on-prem, in another cloud, or air-gapped | xpander |
Employees, not only developers, need a governed way to use agents | xpander |
Using both
Agents you keep on AgentCore can send their model calls to the xpander AI Gateway with an xpander key. Then the same model access, budgets and logs apply to those calls.
Frequently asked questions
Is xpander an alternative to Amazon Bedrock AgentCore?
For companies that want a finished control plane rather than services to assemble, yes. AgentCore gives developers managed services to build their own agent stack on AWS. xpander runs and governs Claude Code, Codex, OpenCode and agents built in plain language, in xpander's cloud, in your Kubernetes or air-gapped.
Can xpander run on AWS?
Yes. xpander installs on Amazon EKS and calls Amazon Bedrock models through its AI Gateway, keyless with EKS Pod Identity. On an air-gapped install, Bedrock is reachable only over a private VPC endpoint. Otherwise, models run in your network, for example on your own GPUs.
Does AgentCore run air-gapped or on-premises?
AWS documents AgentCore in AWS Regions, including AWS GovCloud (US-West), with VPC connectivity. It does not document a self-hosted or air-gapped version. On xpander, air-gapped installs run every component in your network, with no connection to xpander cloud.
Which one governs Claude Code and Codex?
In xpander, Claude Code, Codex and OpenCode run as governed agents: model calls go through the AI Gateway, tool calls run as xpander skills with approvals, and budgets apply. AWS has published how to host coding agents yourself on AgentCore Runtime, including Claude Code with security controls (November 6, 2026). Your team builds and runs that setup.
Facts checked 18 Nov 2026.
Sources
xpander: Deployment options, AI Gateway, Budgets, Approvals, Audit log, Pricing
