Claude in Your Coding Agent: Anthropic's API, Amazon Bedrock, or Any Gateway

Anthropic publishes its own API, the Messages API: POST /v1/messages, a key, and a version header. Amazon Bedrock offers Claude on an endpoint that speaks the same format, and so do many gateways and proxies. So the question Atlarix asks is not "which company" but "which format". A provider you add can now speak either OpenAI's format or Anthropic's, and the agent talks to each in its own terms, through Anthropic's own SDK for the second.

This is the direct route, and it needs nothing custom. Create a key in the Anthropic Console, then:

• Desktop app: Settings → Providers → Anthropic, and paste the key.
• CLI: /connect → Anthropic.
• VS Code, Cursor, Kiro, Windsurf: the account panel → Connect a provider… → Anthropic.

The key is checked with Anthropic before it is saved, Claude's models appear in the model picker, and your requests go from your computer to Anthropic, not through Atlarix. Keys and providers are shared, so connecting once in any of these is enough.

Bedrock's Anthropic-compatible endpoint is https://bedrock-runtime.<region>.amazonaws.com/anthropic/v1, for example us-east-1. Add it as a provider of your own:

• Desktop app: Settings → Providers → add a custom provider, and choose API format: Anthropic.
• CLI: /connect → Other: your own endpoint… → Anthropic.
• VS Code panel: add a provider → Your own endpoint → API format Anthropic.

Give it a name, paste the base URL up to /v1, and use a Bedrock API key as the key (Bedrock's console issues them). Bedrock's endpoint does not list its models, so Atlarix asks for the model id: Bedrock's own id, such as us.anthropic.claude-haiku-4-5-20251001-v1:0. The us. or global. prefix picks a cross-region profile.

Two things on the AWS side, before any of this works. The first time an account uses Anthropic's models on Bedrock, AWS asks for a short use-case form in the Bedrock console. And AWS decides which Claude models each account may call: a new account may see "not available for this account" for the newest ones until AWS enables them.

A company proxy, an LLM gateway, a self-hosted router: if it serves POST /v1/messages, add it the same way as Bedrock, with its own base URL up to /v1 and its own key. When it serves a model list at /v1/models, Atlarix shows those models; otherwise you type the model id once.

Anthropic's API takes the key in an x-api-key header. Bedrock and many gateways take a Bearer token instead. When you add the provider, Atlarix asks the endpoint: it tries x-api-key, then Bearer, and remembers the one that worked. For an endpoint with no model list, it asks with an empty request that costs nothing. A key the endpoint refuses is never saved, and the provider's own error is what you see.

Atlarix recognises the Claude model behind the name, including Bedrock's longer ids, and uses its catalog entry for the rest:

• Thinking and `/effort`: the newest Claude models think adaptively, and /effort sets how hard (low through max, as the model lists them). Older models that take a thinking budget get three budget sizes.
• Prompt caching: the stable part of the prompt is marked for Claude's cache, so later messages in a chat read it instead of paying for it again.
• Tools, plans, approvals: every mode works, Build, Explore, Plan, Debug and Review, with edits and commands waiting for your yes.

One difference: Anthropic's own server-side web search is used only on Anthropic's API. Through Bedrock or a gateway, the agent uses its own web search tool instead.

Atlarix adds nothing to your own provider's bill. Anthropic bills its key at its own prices; Bedrock bills through your AWS account at AWS's rates for that model; a gateway bills however your company set it up. The chat's usage panel shows what the session has cost on your key.