@stratusagent/provider-anthropic

The Claude provider for Stratus Agent, built on the official @anthropic-ai/sdk.

• Multi-turn tool calling — advertises kernel tools with wire-safe names, parses tool_use blocks, and replays results as tool_result blocks.
• Persona and memory — the agent's identity and long-term memory are rendered as the system prompt, so an agent is the same person on every provider.
• Adaptive thinking, handled correctly — claude-opus-5-5 (the default) thinks adaptively; the thinking blocks that precede tool calls are persisted in session metadata and replayed verbatim, surviving tool waits, provider restarts, and resuming a session in another process. Use redactAnthropicRawTurns(session) before showing a session to people — replay state is never meant to be displayed.
• Preserved thinking — Claude Opus 5.5, Fable 5.1 and Sonnet 5.5 bind each thinking block to the conversation before it, and for accounts created on or after 2026-08-31 an edited history is a 400. On those models the provider keeps memory in the system block (which changes only when a memory does) instead of the tail, and asks the API to drop a mismatched block rather than refuse the request (thinking-binding-controls-2026-08-01, drop_block). They also always think, so thinking: 'disabled' is refused for them when the provider is created.
• Auth — an Anthropic API key (apiKey), or an OAuth bearer token (authToken). Note: Claude Pro/Max setup tokens are only honored by Anthropic inside the Claude Code harness, so they do not work against the raw Messages API this provider calls — for subscription-billed runs use @stratusagent/provider-claude-code, which the Stratus CLI selects automatically for subscription sign-ins.

Usage

import { createAnthropicProvider } from '@stratusagent/provider-anthropic';

const provider = createAnthropicProvider({
  apiKey: process.env.ANTHROPIC_API_KEY,
  model: 'claude-opus-5-5',                  // default
});

Pass the provider to AgentRunner from @stratusagent/core and it handles the loop: tools are executed locally and results are fed back to Claude until the model answers in plain text.

Options: model, maxTokens (default 16000), systemPrompt, baseUrl, thinking: 'disabled', and an injectable fetch for tests.

A turn that hits maxTokens before the model finished is refused, not returned: the reply would be a fragment, and a fragment delivered as an answer reads exactly like a complete one. The error names the cap that was in force. Raising maxTokens past roughly 20k only works on the streaming path — the Anthropic SDK refuses a non-streaming request whose cap puts its estimated duration past ten minutes. stratus serve streams; a host calling generate with no onDelta does not.

Because this adapter accepts any model name and any baseUrl, no default is right for every endpoint: one whose own ceiling is below 16000 would have every request refused before generating. Operators set maxTokens in the config file for that case.

Most users won't wire this directly — @stratusagent/cli sets it up from a menu: npm i -g @stratusagent/cli && stratus setup.