Model Providers

Agentwerke selects its model provider through the Anthropic configuration section. The section name remains Anthropic for compatibility, but it can select Anthropic, OpenAI-compatible endpoints, LiteLLM, Azure OpenAI through LiteLLM or OpenAI-compatible routing, the mock provider, or automatic behavior.

Provider choices

Provider	Use when
mock	You need deterministic, tokenless demos, local development, or CI.
anthropic	You are using Anthropic directly.
openai	You are using an OpenAI Chat Completions-compatible endpoint.
litellm	You route model traffic through LiteLLM or another compatible proxy.
auto	You want Agentwerke to infer behavior from configured credentials.

When no usable provider is configured, Agentwerke uses a safe null client and agent steps report that no model is configured.

Configure a mock provider

Use this for local demos and tests:

Anthropic__Provider=mock

Mock runs should produce deterministic output and zero model cost.

Configure a real provider

Set the provider, API key, model, and limits through environment variables, secret stores, appsettings, Helm values, or the Admin Settings page.

Anthropic__Provider=anthropic
Anthropic__ApiKey=<secret>
Anthropic__Model=claude-sonnet-4-6
Anthropic__MaxTokens=4096
Anthropic__MaxToolIterations=10

For an OpenAI-compatible or LiteLLM endpoint:

Anthropic__Provider=litellm
Anthropic__ApiBaseUrl=http://litellm:4000/v1
Anthropic__ApiKey=<secret>
Anthropic__Model=<model-name>

Configure run budgets

Use budgets to prevent runaway agent loops:

Anthropic__MaxRunCostUsd=2.00
Anthropic__MaxRunTokens=100000

Set either value to 0 for unlimited. When a run reaches its budget, Agentwerke stops further model calls with a budget_exceeded status.

Rotate model secrets

Administrators can rotate supported model secrets from Settings. Secret inputs are write-only, the browser clears the value after save, and the API returns only configured/missing status, source, and fingerprint.

For production, prefer deployment-managed secrets over local file-backed settings writes:

Settings__AllowLocalSecretWrites=false

Readiness checks

Use the Settings readiness check for model providers before running production workflows. The check validates local configuration completeness and writes a redacted audit record. It should not mutate external systems.