Anthropic

Anthropic is a first-party model provider — it produces the Claude family of models and serves them from its own API. It is the zero-config default for the claude llm tool, and a first-class model provider in its own right: once connected, operator lists its available models live so delegators can pick one.

Model provider ≠ llm tool. Anthropic (the provider) serves the models; Claude Code (the llm tool) is the CLI that drives a coding session. A delegator pairs a tool with a provider’s model.

Connect

Operator references your key by env-var name — it never stores the secret. Set the standard Anthropic key and operator can probe the provider:

export ANTHROPIC_API_KEY="sk-ant-..."

A provider is connected when its /models probe succeeds (key present + endpoint reachable). In the Model Providers view (web /#/model-providers or the VS Code section) Anthropic then shows ● connected with its live model list.

Listing models

Operator probes https://api.anthropic.com/v1/models and stays agnostic to which models exist — it lists whatever the API returns rather than hardcoding names:

GET /api/v1/model-servers/kinds/anthropic-api/models   # { reachable, models[], error? }

Use from a delegator

Pick a connected provider + one of its live models in the Create Delegator form, or declare it directly. model is a live id from the list above:

[[delegators]]
name = "claude-opus"
llm_tool = "claude"
model = "claude-opus-4-20250514"
# model_server omitted → resolves to the implicit anthropic-api default

Custom endpoint / proxy

To route Claude through a proxy or bridge (e.g. a local model behind claude-code-router), declare an anthropic-api server with an explicit base_url; that base URL is then also injected at spawn (ANTHROPIC_BASE_URL). The probe default (https://api.anthropic.com) is probe-only and never alters the agent’s launch path on its own.