OpenRouter

The openrouter provider reaches hundreds of models from many vendors (Anthropic, OpenAI, Google, Meta, Mistral, DeepSeek, Qwen and more) through one key. OpenRouter serves the OpenAI API, so the planner uses the official OpenAI Java SDK pointed at https://openrouter.ai/api/v1.

Set up

1. Create a key at openrouter.ai and add credit.

2. Export it with a model name, using OpenRouter's own model id:

   export OPENROUTER_API_KEY=sk-or-v1-...
   export NEXORA_LLM_MODEL=meta-llama/llama-4-maverick

OpenRouter model ids are vendor/model, and they write versions with dots: Claude Opus 5.5 is anthropic/claude-opus-5.5, not claude-opus-5-5. List them, with the parameters each supports:

curl -s https://openrouter.ai/api/v1/models | jq -r '.data[] | select(.supported_parameters | index("structured_outputs")) | .id'

Choose a model whose supported_parameters include structured_outputs.

Settings

Variable	Default	Notes
OPENROUTER_API_KEY		Or NEXORA_LLM_API_KEY_FILE
NEXORA_LLM_MODEL	required	An OpenRouter model id
NEXORA_LLM_MAX_TOKENS	16000	See credit below; often worth lowering
NEXORA_LLM_TEMPERATURE	the model's default	0 to 2
NEXORA_LLM_EFFORT		Not accepted; the plugin refuses to start with it set

The overview lists the settings every provider shares.

How it behaves

• Structured output. The plan is requested with a strict json_schema response format, which OpenRouter passes to the model. Models that ignore it are still caught by the plan checks.
• Token limit. Sent as max_tokens.
• Routing. OpenRouter picks which upstream host serves the model. Latency varies by model and host; in testing it ranged from 7 to 19 seconds for the same five-step goal.

Credit and max_tokens

OpenRouter reserves credit for the full max_tokens before it answers. On a key with little credit, the default 16000 can be refused outright with HTTP 402, which the planner reports as REJECTED:

reason=REJECTED detail=OpenRouter API returned HTTP 402: This request requires more credits, or fewer max_tokens.
You requested up to 16000 tokens, but can only afford 2000.

A plan answer itself needs only a few hundred tokens, so lowering NEXORA_LLM_MAX_TOKENS fixes it. Thinking models, Gemini Flash for example, also spend part of the limit on reasoning: give them a few thousand tokens, or the answer comes back TRUNCATED.

When to use OpenRouter

• Comparing models. Change NEXORA_LLM_MODEL and rerun to compare plan quality, latency and cost across vendors, with one key and no code change.
• Vendors without a Java SDK. Meta Llama, Mistral, DeepSeek, Qwen and others.
• One bill. Many models behind one account and one key.

For production traffic to a single vendor, the native provider (Anthropic, OpenAI, Gemini) removes a hop and lets you use that vendor's own features.

Verified

Tested against the real OpenRouter API with a five-step goal ("validate and fraud-check the order, reserve stock, charge it, email a receipt that references the charge"). All four models planned every step correctly, with the receipt's chargeId bound to the charge step's output:

Model	Result	Time
meta-llama/llama-4-maverick	All 5 steps	19 s
openai/gpt-6.1-terra	All 5 steps	7 s
google/gemini-3.8-flash	All 5 steps, at max_tokens 9000	13 s
anthropic/claude-opus-5.5	All 5 steps, at max_tokens 1700	12 s

The low-credit 402, the TRUNCATED answer and an invalid model id all fell back to the rule-based plan as expected.

Troubleshooting

You see	Do this
REJECTED ... HTTP 402 "requires more credits"	Lower NEXORA_LLM_MAX_TOKENS, or add credit
REJECTED ... HTTP 400 "is not a valid model ID"	Use the exact id from the models list, with dots, for example anthropic/claude-opus-5.5
TRUNCATED	Raise NEXORA_LLM_MAX_TOKENS; thinking models need a few thousand
INVALID_PLAN often with one model	That model follows the schema poorly; choose one with structured_outputs