AeionClaw + Anthropic Claude

Configure Anthropic Claude (Opus 4.6, Sonnet 4.6, Haiku 4.5) as AeionClaw's LLM provider. Claude excels at tool use, instruction-following, and high-stakes reasoning — making it the recommended primary for Claw's tool-dispatch-heavy workflows. Per-tenant API key, native tool-use/function-calling and vision support, no-training endpoint, transparent cost attribution.

Anthropic Model Matrix

`

Setup Walkthrough

```bash

Prompt Caching — Available, Not Automatic

`

Tool-Use Reliability — Why Claude for Tool Dispatch

`

Model Choice by Workload

`

FAQ

Any configured text provider works. Claude is recommended for tool-use-heavy workflows (most Claw use cases are tool-use-heavy). For pure conversation without tool dispatch, GPT / Gemini are competitive.

Yes — the model is a per-request parameter (`model` field on the chat completion call), so callers can pick Haiku for a cheap classification and Opus for a deep analysis. There's no persona-to-model assignment step in the AI Providers admin screen today; that routing logic lives in whatever code is issuing the request.

Anthropic's standard API endpoint does not use your data to train models — that's the default for all API customers, not something AeionClaw configures separately.

200K tokens for Sonnet 4.6, Opus 4.6, and Haiku 4.5. Input cost is per-token, so a full 200K-token call costs $15 (Opus) at input pricing alone — most Claw use cases run 10-50K tokens per call.

Not today. Anthropic's `cache_control` prefix caching is available and can cut input cost up to 90% on repeated prefixes, but the Anthropic adapter doesn't set it automatically — it would need to be added by whoever's issuing the request (e.g. a custom skill), via the Anthropic-specific passthrough options on the chat completion call.

Anthropic's extended-thinking parameter (`thinking: { type: "enabled", budget_tokens: N }`) can be passed through on a chat completion call via Anthropic-specific provider options — AeionClaw's adapter forwards arbitrary Anthropic request fields, it doesn't special-case this one.

A hard rate-limit response from Anthropic surfaces as a failed call today — there's no built-in queuing or automatic fallback to a different model/provider. If you need resilience against rate limits, configure a second provider in AI Providers and switch manually, or throttle call volume upstream.

Not currently — AeionClaw's Anthropic adapter talks to Anthropic's direct API only (`api.anthropic.com`). AWS Bedrock and Google Vertex AI routes for Claude aren't wired up.

Anthropic Claude — recommended primary.