LLM Providers

Supported providers

A fourth option, Platform, is available when connected to the Gamut platform. It uses managed credentials and requires no separate API key.

Switching providers

Select the active provider in Settings > LLM Provider. Only one provider is active at a time. Running agents keep using the previous provider until restarted; the UI shows a notice about this when switching.

API key management

API keys can be configured in two ways:

1. Settings UI: Enter the key in Settings > LLM Provider under the credentials section. Keys saved this way are stored locally in settings.json with file permissions restricted to the current user (mode 0600).
2. Environment variables: Set the matching variable before starting Gamut. If both exist, the saved key takes precedence.

A badge shows the current key's source ("Using saved setting" or "Using environment variable"). Remove a saved key to fall back to the environment variable, or save a new key to override it.

Key validation

When a key is entered in the Settings UI, Gamut validates it with a minimal API call (a single-token request to Claude Haiku) and saves it only if validation succeeds. This catches expired keys, incorrect prefixes, and insufficient permissions.

Anthropic (primary)

The Anthropic provider sends requests directly to the Anthropic API at https://api.anthropic.com.

Environment variable: ANTHROPIC_API_KEY

When a container starts, the key is injected via the ANTHROPIC_API_KEY environment variable so the Claude Code process inside the container can authenticate.

OpenRouter

OpenRouter provides an Anthropic-compatible endpoint that routes requests to multiple model providers. Useful for a single API key covering Claude alongside other models, or for routing traffic through OpenRouter for billing or quota reasons.

Environment variable: OPENROUTER_API_KEY

When a container starts, three environment variables are injected:

• ANTHROPIC_API_KEY is set to an empty string (prevents the SDK from sending the x-api-key header).
• ANTHROPIC_BASE_URL is set to https://openrouter.ai/api.
• ANTHROPIC_AUTH_TOKEN carries the OpenRouter key via the Authorization: Bearer header.

Under OpenRouter, the agent default is sonnet (Claude Sonnet 5) rather than opus, as a cost-conscious default.

AWS Bedrock

AWS Bedrock provides Claude inference through Amazon's managed AI infrastructure. This is the recommended option for enterprises that route all AI traffic through their AWS account for compliance, cost allocation, or network isolation.

Authentication methods

Simple auth (Bearer Token):

• Set the AWS_BEARER_TOKEN_BEDROCK environment variable or enter the token in Settings > LLM Provider.
• The simplest option when a Bedrock-specific API key is available.

Full AWS credentials (Access Key + Secret):

• Set AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY as environment variables, or enter them in Settings > LLM Provider.
• Configure the AWS region (defaults to us-east-1) via the AWS_REGION environment variable or in settings.

If neither method is configured, Bedrock falls back to the default AWS credential chain (~/.aws/credentials or an instance profile).

When a container starts with Bedrock active, it receives CLAUDE_CODE_USE_BEDROCK=1 to put the Claude Code SDK in Bedrock mode. ANTHROPIC_API_KEY is explicitly cleared so the container cannot fall back to the direct Anthropic API. Bedrock uses cross-region model IDs (the us.anthropic. prefix); the catalog lists these when Bedrock is active.

Models

Each provider exposes a dynamic model catalog in Settings > LLM Provider. Models can be selected two ways:

• Family aliases (opus, sonnet, haiku): Resolve to the latest version of that family, so new releases apply without a settings change.
• Version pinning: Pin an exact version from the catalog when a workflow depends on stable model behavior.

The catalog covers the Claude Opus, Sonnet, and Haiku families, Fable 5 at all effort levels, and GPT models (gpt-5.4, gpt-5.5) through OpenRouter and the Platform provider.

Defaults

The Default Model applies to new agent sessions when no per-message model is selected. The Summarizer Model handles lightweight work such as session name generation and API key validation.

Per-message model selector

The message composer can switch between the three Claude families on a per-message basis:

• Opus: Most capable; best for complex, multi-step tasks.
• Sonnet: Balanced speed and capability.
• Haiku: Fastest and most affordable.

The family selected in the composer applies to that message and persists for subsequent messages in the same session until changed.

Tool search

The Tool Search toggle in Settings > LLM Provider controls whether agent containers load tool definitions on demand via a meta-tool or all at once upfront. Enabled by default, it saves roughly 15,000 to 20,000 context tokens per turn across Gamut's 60+ tool surface. Tool search requires Sonnet 4+ or Opus 4+ and is ignored on Haiku.

Disable this only when debugging tool-loading behavior.

Environment variables reference