Model providers

{/* Imported from CoderHelm/coderhelm-platform/docs/self-host/model-providers.md at 96a2a3bafd5e. Edit it there. */} CoderHelm uses Claude for every run, review, plan chat and log analysis. You pick how it is reached with models.provider in the config.

Anthropic API (default)Copy Anchor Link

"models": { "provider": "anthropic" }

Each team adds its own Anthropic API key in Settings → Model Provider, and picks its models there. Calls go from your Lambda functions straight to api.anthropic.com; the key is stored on the team's settings item in your DynamoDB table (encrypted with your KMS key) and is never sent anywhere else.

Amazon BedrockCopy Anchor Link

"models": {
  "provider": "bedrock",
  "bedrockInferencePrefix": "us."
}

The gateway, streaming, worker and log analyzer functions call Bedrock with their own IAM roles. No API key exists anywhere; usage and cost land on your AWS bill, and data stays in AWS.

Before you deploy:

1. In the Bedrock console of your region, open Model access and enable the Claude models you will use (by default Claude Opus 5.5 and Claude Sonnet 5.5). Some models ask you to accept terms first.
2. Pick the inference profile prefix for your region: us. for US regions, eu. for Europe, apac. for Asia Pacific, global. for the global profiles, or "" to call a model in your region directly.

How model names map to Bedrock ids. Teams and the config use Anthropic's model names (claude-opus-5-5). CoderHelm turns them into Bedrock ids:

• an entry in models.bedrockModelIds wins, for example an application inference profile ARN:

  Copy Text

  "bedrockModelIds": {
    "claude-opus-5-5": "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123"
  }

• a name that already is a Bedrock id (contains anthropic.) is used as it is;
• otherwise it becomes <prefix>anthropic.<name>, so claude-sonnet-5-5 with us. is us.anthropic.claude-sonnet-5-5. Models whose Bedrock id carries a date (Claude Haiku 4.5) are mapped to the dated id for you.

What the roles may call: bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on Anthropic foundation models and your account's inference profiles, nothing else.

With Bedrock, the Settings → Model Provider page says the install uses Bedrock and no key is needed; teams can still choose which models they use.

Which to chooseCopy Anchor Link

EmbeddingsCopy Anchor Link

Agent memory uses Amazon Titan Text Embeddings V2 on Bedrock in your account with either provider. Enable it in Model access too; without it, memory search is off and runs still work.

Usage recordsCopy Anchor Link

Every model call is counted per team in your analytics table, with either provider. The daily reconciliation against Anthropic's usage report only runs when you add an Anthropic admin key secret (coderhelm/<stage>/anthropic-admin-keys); self-hosted installs normally leave it out.