Alternative AI Providers¶

By default, AI Sessions connects to Claude through Anthropic's API using your Anthropic API key. You can also route Claude requests through Amazon Bedrock or Google Cloud Vertex AI.

This is useful when:

Your organization already has an AWS or GCP agreement that includes Claude access
You need to keep AI traffic within specific regions for compliance
You want to consolidate billing through your existing cloud account

This page is about running Claude through your cloud agreement. To run user sessions on other vendors' models (OpenAI, Google Gemini, OpenRouter), see AI Providers.

Two configuration formats¶

There are two ways to point a session at Bedrock or Vertex:

Structured provider blocks — a top-level bedrock: or vertex: block. On a direct SessionRequest (the org-scoped API), the block is validated and applied: the runner translates it into the correct environment variables for the Claude subprocess.
Manual environment variables — set CLAUDE_CODE_USE_BEDROCK / CLAUDE_CODE_USE_VERTEX and the corresponding cloud-provider variables under the profile's environment: map. This is the original mechanism, and it is the one to use on ai_agent Hive records today.

Availability on ai_agent Hive records

bedrock: and vertex: blocks on ai_agent Hive records are accepted and validated by the schema, but sessions launched from a Hive record (D&R rules, UI actions, ai start-session --definition) do not apply them yet. For those launches, use the manual environment-variable mode described below, which requires anthropic_secret to be set (a placeholder for Bedrock). The structured blocks work today on direct SessionRequest API calls.

Pick exactly one credential source per session: anthropic_secret, the bedrock: block, or the vertex: block. They are mutually exclusive — a session cannot mix providers.

Amazon Bedrock¶

Amazon Bedrock provides access to Claude models through AWS infrastructure.

Required AWS setup¶

IAM permissions¶

The AWS credentials must have permissions to invoke Claude models via Bedrock. At minimum:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream"
      ],
      "Resource": "arn:aws:bedrock:*::foundation-model/anthropic.*"
    }
  ]
}

You must also ensure that the Claude models you intend to use are enabled in your Bedrock console for the selected region.

Model IDs¶

Bedrock model IDs differ from standard Anthropic model IDs — they include a region prefix and version suffix. Set the model field on the profile to one of:

us.anthropic.claude-sonnet-4-5-20250929-v1:0
us.anthropic.claude-haiku-4-5-20251001-v1:0
eu.anthropic.claude-sonnet-4-5-20250929-v1:0
ap.anthropic.claude-sonnet-4-5-20250929-v1:0

The general format is <region-prefix>.anthropic.<model-name>-v<version>:<minor>. The region prefix (us, eu, ap, …) should correspond to your AWS region. Available IDs are listed in the Bedrock model IDs documentation.

Configuration via the `bedrock:` block¶

The bedrock block lives at the top of an ai_agent Hive record, alongside prompt. All credential fields end with _secret and accept either a literal value or a hive://secret/<name> reference.

Warning

Records with a bedrock: block validate and store, but launches from Hive records do not apply the block yet — use the manual environment-variable mode for record-based launches today. The equivalent direct SessionRequest form works now.

ai_agent:
  bedrock-investigator:
    data:
      prompt: "Investigate this detection..."
      lc_api_key_secret: hive://secret/lc-api-key
      model: us.anthropic.claude-sonnet-4-5-20250929-v1:0

      bedrock:
        region: us-east-1
        access_key_id_secret: hive://secret/aws-access-key-id
        secret_access_key_secret: hive://secret/aws-secret-access-key
        # Optional — only when using STS / SSO temporary credentials:
        session_token_secret: hive://secret/aws-session-token
    usr_mtd:
      enabled: true

`bedrock` field reference¶

Field	Required	Description
`region`	Yes	AWS region where Bedrock is available (for example `us-east-1`, `us-west-2`, `eu-central-1`, `ap-southeast-2`). Sets `AWS_REGION` on the runner.
`access_key_id_secret`	Conditional	AWS access key ID, or a `hive://secret/<name>` reference. Sets `AWS_ACCESS_KEY_ID`. Must be paired with `secret_access_key_secret`.
`secret_access_key_secret`	Conditional	AWS secret access key, or a `hive://secret/<name>` reference. Sets `AWS_SECRET_ACCESS_KEY`. Must be paired with `access_key_id_secret`.
`session_token_secret`	No	Temporary session token from STS or SSO, or a `hive://secret/<name>` reference. Sets `AWS_SESSION_TOKEN`. Requires the access-key pair.
`bearer_token_secret`	Conditional	Bedrock API bearer token, or a `hive://secret/<name>` reference. Sets `AWS_BEARER_TOKEN_BEDROCK`. Used as an alternative to the access-key pair.

You must supply either (access_key_id_secret + secret_access_key_secret) or bearer_token_secret. The schema rejects records that set neither, and rejects setting only one of the access-key pair.

When a session launches with a bedrock provider block, the runner sets CLAUDE_CODE_USE_BEDROCK=1 automatically — you do not need to add it yourself.

Direct `SessionRequest` (API and integrations)¶

The same provider block is exposed on the AI Sessions SessionRequest type used by the org-scoped API and by integrations that build sessions programmatically. The field names drop the _secret suffix because the values are already-resolved literals at that point:

{
  "prompt": "Investigate this detection...",
  "bedrock": {
    "region": "us-east-1",
    "access_key_id": "AKIA…",
    "secret_access_key": "…",
    "session_token": "…",
    "bearer_token": "…"
  },
  "profile": {
    "model": "us.anthropic.claude-sonnet-4-5-20250929-v1:0",
    "allowed_tools": ["Bash", "Read", "Grep", "Glob", "WebFetch"]
  }
}

Validation enforces exactly one of anthropic_key, bedrock, or vertex per request, plus the same per-block rules listed above.

Configuration via environment variables (manual mode)¶

The original mechanism — setting AWS variables under the profile's environment: map — still works. The runner forwards every entry of environment: to the Claude subprocess as-is, so the cloud-provider variables get picked up there.

This is the mode to use for sessions launched from ai_agent Hive records today (D&R rules, UI actions, ai start-session --definition).

ai_agent:
  bedrock-investigator:
    data:
      prompt: "Investigate this detection..."
      anthropic_secret: hive://secret/anthropic-key  # placeholder, see note below
      lc_api_key_secret: hive://secret/lc-api-key
      model: us.anthropic.claude-sonnet-4-5-20250929-v1:0
      environment:
        CLAUDE_CODE_USE_BEDROCK: "1"
        AWS_ACCESS_KEY_ID: hive://secret/aws-access-key-id
        AWS_SECRET_ACCESS_KEY: hive://secret/aws-secret-access-key
        AWS_REGION: us-east-1
    usr_mtd:
      enabled: true

Variable	Description
`CLAUDE_CODE_USE_BEDROCK`	Must be set to `1` to enable Bedrock.
`AWS_ACCESS_KEY_ID`	AWS access key ID with Bedrock permissions.
`AWS_SECRET_ACCESS_KEY`	AWS secret access key.
`AWS_REGION`	AWS region, matching the model ID's region prefix.
`AWS_SESSION_TOKEN`	(optional) STS/SSO temporary session token.
`AWS_BEARER_TOKEN_BEDROCK`	(optional) Bedrock API bearer token, alternative to access keys.

When using the manual environment-variable form, the schema still requires anthropic_secret to be set on the record. Point it at a hive://secret/<name> containing any non-empty placeholder — the runner ignores it once CLAUDE_CODE_USE_BEDROCK=1 is in the environment.

Google Cloud Vertex AI¶

Google Cloud Vertex AI provides access to Claude models through GCP. Authentication uses a service-account JSON key with the appropriate Vertex AI permissions.

Required GCP setup¶

Enable the Vertex AI API in your project.
Subscribe to the Claude models you intend to use in Vertex AI Model Garden.
Create a service account with at least the roles/aiplatform.user role (or a custom role permitting aiplatform.endpoints.predict).
Generate and download a JSON key for that service account.

Model IDs and region¶

Vertex uses Claude model IDs in the form Anthropic ships them on the platform — typically claude-<model>@<version>, for example claude-sonnet-4-5@20250929. Confirm the IDs available in your project against the Vertex Model Garden listings.

The region you set must be one that Anthropic publishes models to (commonly global, us-east5, or europe-west1). Cross-check with the Anthropic on Vertex AI documentation for current region availability.

Configuration via the `vertex:` block¶

Warning

Records with a vertex: block validate and store, but launches from Hive records do not apply the block yet. Because the manual environment-variable mode cannot carry the service-account JSON (see below), Claude-on-Vertex is currently only usable through the direct SessionRequest form, not from ai_agent Hive records.

ai_agent:
  vertex-investigator:
    data:
      prompt: "Investigate this detection..."
      lc_api_key_secret: hive://secret/lc-api-key
      model: claude-sonnet-4-5@20250929

      vertex:
        project_id: my-gcp-project
        region: us-east5
        service_account_json_secret: hive://secret/vertex-service-account
    usr_mtd:
      enabled: true

`vertex` field reference¶

Field	Required	Description
`project_id`	Yes	GCP project ID hosting the Vertex AI subscription. Sets `ANTHROPIC_VERTEX_PROJECT_ID`.
`region`	Yes	Vertex region (`global`, `us-east5`, `europe-west1`, …). Sets `CLOUD_ML_REGION`.
`service_account_json_secret`	Yes	Full service-account JSON key contents, or a `hive://secret/<name>` reference to a secret holding the JSON.

The runner writes the resolved service-account JSON to a per-session temporary file (mode 0600, removed when the process exits) and points GOOGLE_APPLICATION_CREDENTIALS at it. It also sets CLAUDE_CODE_USE_VERTEX=1 automatically.

Store the entire service-account JSON in a single Hive Secret and reference it via hive://secret/<name>. The JSON contains a private key — never paste it as a literal into a record or D&R rule.

Direct `SessionRequest` (API and integrations)¶

{
  "prompt": "Investigate this detection...",
  "vertex": {
    "project_id": "my-gcp-project",
    "region": "us-east5",
    "service_account_json": "{\"type\":\"service_account\",\"project_id\":\"…\",\"private_key\":\"…\"}"
  },
  "profile": {
    "model": "claude-sonnet-4-5@20250929"
  }
}

service_account_json is the literal JSON document for the service-account key — typically the entire contents of the file you downloaded from GCP, embedded as a JSON string.

Configuration via environment variables (manual mode)¶

If you must configure Vertex through the profile environment: map instead of the structured vertex: block, set the variables the runner would otherwise set on your behalf. Note that you cannot inline the service-account JSON as an environment variable — you have to mount it as a file at a known path inside the runner image and point GOOGLE_APPLICATION_CREDENTIALS at that path. Most users do not have a way to do that, which is why the structured vertex: block on a direct SessionRequest is the supported path.

Variable	Description
`CLAUDE_CODE_USE_VERTEX`	Must be set to `1` to enable Vertex.
`ANTHROPIC_VERTEX_PROJECT_ID`	GCP project ID for the Vertex subscription.
`CLOUD_ML_REGION`	Vertex region.
`GOOGLE_APPLICATION_CREDENTIALS`	Filesystem path to the service-account JSON key.

Storing credentials securely¶

Always store cloud-provider credentials in Hive Secrets and reference them via hive://secret/<name>. Treat the Vertex service-account JSON as a single secret (don't try to split it into multiple fields). For Bedrock, store the access key, secret key, and any session token as separate secrets.

The endpoint resolves hive://secret/<name> references just before sending the request to AI Sessions, so secret contents never appear in D&R rules, argv, or session metadata.

Notes¶

On a direct SessionRequest with a bedrock or vertex block, you do not supply anthropic_key. The ai_agent schema likewise accepts a record with bedrock: or vertex: and no anthropic_secret, but since record-based launches do not apply the blocks yet, record-based Bedrock launches use the manual environment-variable mode, which requires a placeholder anthropic_secret.
Claude model availability varies by AWS region and Vertex region. Check the Bedrock model availability page and Vertex AI Model Garden before picking a region.
Billing for Claude usage goes through your AWS or GCP account when using these providers, not through Anthropic directly.
The provider you choose only affects the Claude API path. LimaCharlie data, MCP servers, the LC CLI, tool execution, and session storage are unaffected.

Alternative AI Providers¶

Two configuration formats¶

Amazon Bedrock¶

Required AWS setup¶

IAM permissions¶

Model IDs¶

Configuration via the bedrock: block¶

bedrock field reference¶

Direct SessionRequest (API and integrations)¶

Configuration via environment variables (manual mode)¶

Google Cloud Vertex AI¶

Required GCP setup¶

Model IDs and region¶

Configuration via the vertex: block¶

vertex field reference¶

Direct SessionRequest (API and integrations)¶

Configuration via environment variables (manual mode)¶

Storing credentials securely¶

Notes¶

Configuration via the `bedrock:` block¶

`bedrock` field reference¶

Direct `SessionRequest` (API and integrations)¶

Configuration via the `vertex:` block¶

`vertex` field reference¶

Direct `SessionRequest` (API and integrations)¶