Anthropic (Claude)

The anthropic provider plans with Claude models through the official Anthropic Java SDK. It is the only provider with a default model, so a key is all you need to start.

Set up

1. Create an API key in the Claude Console.

2. Export it where the engine runs:

   export ANTHROPIC_API_KEY=sk-ant-...

That is enough: with only ANTHROPIC_API_KEY set, the planner selects Anthropic and plans with claude-opus-5-5.

To be explicit, or to pick another model:

export NEXORA_LLM_PROVIDER=anthropic
export NEXORA_LLM_MODEL=claude-sonnet-5-5

Without ANTHROPIC_API_KEY, the SDK looks for credentials itself: ANTHROPIC_AUTH_TOKEN, or a profile saved by ant auth login. Set NEXORA_LLM_PROVIDER=anthropic to use those.

Settings

Variable	Default	Notes
ANTHROPIC_API_KEY		Or NEXORA_LLM_API_KEY_FILE pointing at a file holding it
NEXORA_LLM_MODEL	claude-opus-5-5	For example claude-sonnet-5-5 or claude-haiku-5-5
NEXORA_LLM_EFFORT	the model's default	low, medium, high, xhigh or max. Controls how much the model reasons before answering
NEXORA_LLM_TEMPERATURE		Not accepted: current Claude models reject it, so the plugin refuses to start with it set. Use effort instead
NEXORA_LLM_BASE_URL	Anthropic's API	For a proxy or gateway

The overview lists the settings every provider shares.

How it behaves

• Structured output. The plan is requested as JSON constrained by the plan schema (output_config.format), so the answer always parses.
• Refusal fallback. On claude-opus-5-5, claude-opus-5, claude-sonnet-5-5 and claude-fable-5-1, requests opt into server-side refusal fallback: if the model declines, the API answers with a fallback model instead of a refusal. Other models do not send it.
• Reasoning. Current Claude models think before they answer. Thinking counts toward NEXORA_LLM_MAX_TOKENS, which is why the default is a generous 16000. Lower NEXORA_LLM_EFFORT to make planning faster and cheaper.
• Retries. The SDK retries rate limits, overload and server errors up to NEXORA_LLM_MAX_RETRIES times before the planner falls back.

Choosing a model

Model	Pick it when
claude-opus-5-5 (default)	Goals are long, ambiguous or need many steps wired together
claude-sonnet-5-5	You want most of Opus's planning quality for less cost and latency
claude-haiku-5-5	Goals are short and the capability set is small, and cost matters most

Start with the default, watch the INVALID_PLAN rate in your logs, and move to a smaller model if plans stay correct.

Verified

Tested against the real Anthropic API with claude-opus-5-5:

• A two-step goal planned in about 5 seconds, binding the receipt's chargeId to the charge step's output.
• A five-step goal planned in about 9 seconds, with independent steps run in parallel.
• A goal no capability fits ("Book a flight to Paris") came back empty and fell back to the rule-based plan (INVALID_PLAN).
• An invalid key fell back with REJECTED and the API's 401 message.

Troubleshooting

You see	Do this
REJECTED ... HTTP 401	The key is wrong or revoked. Check ANTHROPIC_API_KEY
REJECTED ... HTTP 404 or 400 naming the model	Check NEXORA_LLM_MODEL against the models your account can use
Plugin fails to start: "temperature is not supported"	Remove NEXORA_LLM_TEMPERATURE; set NEXORA_LLM_EFFORT instead
TRUNCATED	Raise NEXORA_LLM_MAX_TOKENS, or lower NEXORA_LLM_EFFORT so less of it goes to reasoning
UNAVAILABLE ... HTTP 529	The API is overloaded; the SDK already retried. Raise NEXORA_LLM_MAX_RETRIES or try later