Coding agents on AWS Bedrock, Microsoft Foundry and Google Cloud
Published
Summary
Which coding agents run on Amazon Bedrock, Microsoft Foundry or Azure OpenAI, and Google Cloud Vertex AI, how to turn each one on, and what changes from a subscription.
Most coding agents can send their model requests through your own cloud account instead of the agent vendor's. On AWS that means Amazon Bedrock, on Azure it means Microsoft Foundry (which includes Azure OpenAI), and on Google Cloud it means the Agent Platform, which Google used to call Vertex AI. This guide lists which agents support each one, how to turn it on, and what changes when you do.
Vendor docs change often, so every claim about a vendor below links to its own page. We checked them all on November 8, 2026.
Why teams route coding agents through their cloud
Teams usually do this for one of three reasons:
| Reason | What it gets you |
|---|---|
| Billing | Model usage lands on the cloud bill you already pay, and can count toward committed spend with that provider |
| Access control | Engineers sign in with the same IAM, Entra ID, or Google Cloud identities as everything else, and admins revoke access in one place |
| Data terms | Prompts and code go to a model hosted under your existing cloud agreement, instead of a new contract with a model vendor |
The tradeoff is that you pay per token at your cloud provider's rates. There is no flat monthly price, and no usage limit beyond your own quotas and spend limits.
Which coding agents run on Bedrock, Foundry, and Google Cloud
This table covers each agent's own documented setup. "Not documented" means the vendor's pages don't say it works. Ask the vendor before you rely on it. In alphabetical order:
| Agent | Amazon Bedrock | Microsoft Foundry or Azure OpenAI | Google Cloud (Vertex AI) |
|---|---|---|---|
| Claude Code (CLI, IDE, desktop) | Yes (Anthropic) | Yes (Anthropic) | Yes (Anthropic) |
| Claude Code cloud sessions | No. Cloud sessions need a claude.ai sign-in (Anthropic) | No | No |
| Cline | Yes (Cline) | Not documented | Yes (Cline) |
| Codex CLI and IDE extension | Yes, built-in amazon-bedrock provider (OpenAI) | Yes, as a custom Azure provider (OpenAI) | Not documented |
| Codex cloud | No. Codex cloud needs a ChatGPT sign-in (OpenAI) | No | No |
| Cursor | Yes, in the editor (Cursor). Not documented for Cloud Agents | Not documented | Not documented |
| Factory | Yes, in the Droid CLI and desktop app, not the hosted web platform (Factory) | Not documented | Not documented |
| GitHub Copilot | Yes, in Copilot Chat, Copilot CLI and IDEs, in public preview (GitHub). Not documented for the cloud agent | Yes, same surfaces (GitHub) | Not documented. Google AI Studio is listed instead |
| OpenCode | Yes (OpenCode) | Yes (OpenCode) | Yes (OpenCode) |
| Replicas (cloud workspaces) | Yes, for Claude Code (Replicas) | Yes, for Claude Code and Codex (Replicas) | No |
The pattern is clear. Agents that run on your own machine almost all support at least one cloud provider. The vendor-hosted cloud agents mostly don't: Anthropic, OpenAI, Cursor, Factory and GitHub tie their cloud agents to their own subscriptions or plans. If you want an agent that works in the cloud and still bills to your cloud account, check this before you buy.
How to turn it on in each agent
Each agent has its own switch. These are the documented settings:
| Agent | Amazon Bedrock | Microsoft Foundry or Azure OpenAI | Google Cloud |
|---|---|---|---|
| Claude Code | CLAUDE_CODE_USE_BEDROCK=1, or run /setup-bedrock | CLAUDE_CODE_USE_FOUNDRY=1 and ANTHROPIC_FOUNDRY_RESOURCE | CLAUDE_CODE_USE_VERTEX=1, CLOUD_ML_REGION, and ANTHROPIC_VERTEX_PROJECT_ID |
| Codex | model_provider = "amazon-bedrock" in config.toml | A [model_providers.azure] entry with your endpoint and AZURE_OPENAI_API_KEY | Not documented |
| Cursor | An IAM role in the Cursor dashboard, then each user turns on Bedrock under Cursor Settings → Models | Not documented | Not documented |
| GitHub Copilot | An enterprise owner enables custom models, then an admin adds the key | Same as Bedrock | Not documented |
| OpenCode | AWS_PROFILE or AWS_BEARER_TOKEN_BEDROCK, plus AWS_REGION | AZURE_RESOURCE_NAME, with an API key or Entra ID | GOOGLE_CLOUD_PROJECT and GOOGLE_VERTEX_LOCATION |
On every provider you also have to turn on the models in your cloud account first. On Bedrock you submit Anthropic's use case form once per AWS account. On Google Cloud you request access in Model Garden, and Anthropic's docs say approval may take 24 to 48 hours (Anthropic). On Foundry you create a deployment for each model (Anthropic).
What changes compared with a subscription
Moving Claude Code from a Claude subscription to a cloud provider changes a few things. Anthropic documents these:
| Change | Bedrock | Foundry | Google Cloud |
|---|---|---|---|
| Sign-in | AWS credentials, SSO profile, or a Bedrock API key | API key, Entra ID, or a bearer token | Application Default Credentials or a service account |
| Web search tool | Not available (Anthropic) | Available | Available on Opus 4, Sonnet 4, Haiku 4.5 and later models |
| Model check at startup | Yes. Falls back to an older model if the default isn't enabled | No. Requests fail if the default model isn't deployed | Yes. Same fallback as Bedrock |
| Where inference runs | AWS | Azure or Anthropic, depending on the hosting option you pick for the deployment | Google Cloud |
Two things apply to all three. First, pin model versions before you roll out to a team. Without pins, the opus and sonnet aliases resolve to Claude Code's built-in defaults, which may not be enabled in your account. Second, on Bedrock and Google Cloud, Claude Code uses Opus by default if you don't pin a primary model, and Opus costs more per token than Sonnet (Anthropic).
Questions to ask a coding agent vendor
- Does the cloud agent itself run on our Bedrock, Foundry, or Google Cloud account, or only the editor and CLI?
- How does it sign in: a static access key, an IAM role it assumes, or workload identity?
- Where are the credentials stored, and who can read them?
- Can we set the cloud account as the default and keep subscriptions as a fallback, or the other way around?
- Does the vendor add any fee on model usage that goes through our account?
Where Replicas fits
Replicas runs Claude Code, Codex, Cursor, OpenCode and Pi in cloud workspaces, on model accounts you already have. Claude Code can run on Amazon Bedrock or Microsoft Foundry, and Codex can run on Azure OpenAI through Microsoft Foundry. Google Cloud isn't supported yet. Bedrock uses an access key ID, secret access key and region, not an assumed IAM role (Replicas).
An admin decides the order in which accounts are tried, called a fallback chain. A company with committed AWS spend can put Bedrock first, and a team can keep Claude subscriptions first and fall back to Bedrock when a subscription hits its usage limit. Replicas charges for workspace time and doesn't add a fee on model usage. Any organization member who can open a workspace can read the credentials inside it, so use a key scoped to model access only.
For subscriptions and plain API keys, see Cloud coding agents that use your own API key or subscription.
FAQ