Hot AI News

Kimi Code CLI now signs in to Bedrock with your AWS login, no API key

Moonshot AI's open source coding agent v2.7.0 reads standard AWS credentials, including SSO profiles, to reach Kimi K3.1 Pro on Amazon Bedrock without a separate key.

Kimi Code CLI now signs in to Bedrock with your AWS login, no API key
Image: Amazon

Moonshot AI’s open source coding agent, Kimi Code CLI, can now reach models on Amazon Bedrock through the standard AWS credential chain. According to an AWS Machine Learning blog post, v2.7.0 adds a native bedrock provider, so no separate Bedrock API key is needed.

Named profiles and SSO replace a separate key

Before this release, connecting Kimi Code CLI to Bedrock meant defining a generic provider of type openai or openai_responses, setting a regional base_url and supplying a Bedrock API key. AWS describes that key as an extra credential to obtain and protect.

Kimi Code CLI now signs in to Bedrock with your AWS login, no API key
Image: Amazon

The new provider resolves credentials the way the AWS CLI does. It can use the default chain or a named profile selected with AWS_PROFILE, including profiles backed by IAM Identity Center single sign-on. Per AWS, the provider then uses the @aws/bedrock-token-generator package to mint a short-term Bedrock key and renew it when needed. Those keys last up to 12 hours or the length of the AWS session, whichever is shorter. A key set explicitly through api_key or api_key_env still takes precedence.

AWS uses Kimi K3.1 Pro Highspeed as the worked example. It lists streaming, configurable reasoning, image input and client-side tool calling as supported. The sample config.toml points a model alias at the global inference profile global.moonshotai.kimi-k3.1 and declares a 1,000,000-token context size. The setup needs an AWS account with access to the model, an IAM identity allowed to invoke it, Kimi Code CLI 2.7.0 or later and the AWS CLI. The optional aws login flow needs AWS CLI 2.32.0 or later.

The post includes a least-privilege IAM policy that allows bedrock:CallWithBearerToken for short-term tokens only, plus invoke rights scoped to that one inference profile. One statement, for the project/default resource, is required only if you switch to the openai_responses protocol.

Advertisements

Region, capability and status quirks to know

AWS lists several caveats. The provider needs AWS_REGION or AWS_DEFAULT_REGION set in the Kimi Code config.toml or in the shell. It cannot read a Region configured only in ~/.aws/config. With cross-Region profiles, that setting decides which source Region receives requests.

Kimi Code does not auto-detect the capabilities of most Bedrock models, so users must declare them by hand. For K3.1 Pro Highspeed that means always_thinking, thinking, image_in and tool_use, but not audio or video input. At the time of writing, the desktop and web provider screens do not expose every environment field this setup needs, so AWS says to edit config.toml directly.

The provider infers the API protocol and endpoint from the model ID. For Kimi models it picks the OpenAI-compatible Chat Completions path. Users can override that with protocol set to openai, openai_responses or anthropic, and with base_url for cases such as a VPC endpoint or the Mantle endpoints.

The post also warns that the “connected” status in the desktop and web apps does not check credentials, permissions, Region, model access or network path. AWS recommends a non-interactive kimi --prompt request as the real test.

A Chinese lab’s model, sold through an American cloud

The release is small, but it shows how Bedrock is positioned. AWS announced Kimi K3.1 Pro on the platform on Nov. 11, and nine days later it is publishing setup guidance for the lab’s own agent. Bedrock is pay-per-token, and AWS says the global profile is priced 10% below geographic ones for this model. That gives AWS customers a route to Moonshot’s models that stays inside their existing billing and identity setup.

It also lands while Moonshot’s access to American infrastructure is in question. As we reported on Nov. 19, Moonshot reportedly dropped a $150M rental of Nvidia Blackwell Ultra B300 chips in Indonesia. The report came from The Information, citing two anonymous people, and neither company has confirmed it. It followed a new interim final rule from the US Commerce Department’s Bureau of Industry and Security. The rule limits remote access to advanced chips abroad for firms headquartered in China, and it names Indonesia among the countries needing licenses. Serving Kimi through Bedrock does not touch that rule: inference runs on AWS capacity under AWS’s terms. The source does not say whether Moonshot gets any commercial benefit from Bedrock usage, and that is not covered in the announcement.

Advertisements

What the announcement does not establish

The post is AWS’s own documentation, and its statements about security and pricing are the company’s claims. It gives no per-token prices for Kimi K3.1 Pro Highspeed and sends readers to the Bedrock pricing page and the model card for Region and inference-profile support. It makes no performance claims for the CLI itself. The context window and capability flags in the sample config are what the example declares, and AWS tells users to validate tool and image workflows themselves.

For teams already on AWS, the practical change is narrower than the headline suggests. Upgrade to 2.7.0, add a few lines to config.toml, and sign in with an existing profile. AWS still says protecting those credentials and scoping permissions remains the user’s job.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like