Configuration

Complete reference for all Axiom Protocol plugin configuration parameters.

Full Configuration Example

openclaw.plugin.json

{
  "workerUrl": "https://your-worker.workers.dev",
  "apiKey": "sk-or-v1-...",
  "defaultModels": [
    "openai/gpt-5.4-mini",
    "anthropic/claude-haiku-4-5",
    "google/gemini-2.5-flash"
  ],
  "maxIterations": 7,
  "structure": "fugue",
  "dThreshold": 0.05
}

Parameters Reference

Parameter	Type	Required	Default	Description
workerUrl	string	Yes	—	Axiom Protocol consensus worker endpoint. Deploy your own from ACP-PROJECT.
apiKey	string	Yes	env: OPENROUTER_API_KEY	OpenRouter API key. Can be set via environment variable or passed directly.
defaultModels	string[]	No	gpt-5.4-mini, claude-haiku-4-5, gemini-2.5-flash	Default model identifiers. Minimum 2 required for consensus.
maxIterations	number	No	7	Maximum convergence iterations (1–7). 7 iterations = 3.4% remaining disagreement.
structure	enum	No	fugue	Consensus structure: fugue (layered analysis), sonata (thesis→antithesis→synthesis), or concert (leader + ensemble).
dThreshold	number	No	0.05	D-score threshold (0–1). Consensus is reached when D drops below this value.

Worker URL

The workerUrl points to your Axiom Protocol consensus worker. The plugin sends POST requests to {workerUrl}/consensus-iterative.

Deploy the worker from the ACP-PROJECT repository. The worker runs on Cloudflare Workers and handles the iterative consensus algorithm.

Trailing slashes

The plugin automatically strips trailing slashes from the worker URL. Both https://worker.dev and https://worker.dev/ will work correctly.

API Key

The plugin requires an OpenRouter API key to access LLM models. You can provide it in two ways:

Environment variable (recommended): Set OPENROUTER_API_KEY in your environment
Direct config: Pass the key in the apiKey field of the plugin configuration

The key is sent in the x-openrouter-key HTTP header with each consensus request.

Default Models

The defaultModels array specifies which LLMs participate in consensus by default. You need at least 2 models. For best results, use models from different providers to maximize diversity of reasoning.

Recommended models

[
  "openai/gpt-5.4-mini",
  "anthropic/claude-haiku-4-5",
  "google/gemini-2.5-flash"
]

Cross-provider diversity

Models from different providers (OpenAI, Anthropic, Google) have different training data and architectures, making axiom grounding more effective.

Max Iterations

Controls how many convergence rounds the algorithm performs. Each iteration reduces the D-score by the golden ratio (φ ≈ 1.618):

1 iteration: ~61.8% of original disagreement remains
3 iterations: ~23.6% remaining
5 iterations: ~9.0% remaining
7 iterations (default): ~3.4% remaining

Higher iterations give stronger consensus but increase latency and API costs.

Consensus Structure

Structure	Pattern	Best For
fugue	Models build on each other in layers	Complex questions, technical analysis, deep research
sonata	Thesis → antithesis → synthesis	Debates, ethical dilemmas, competing approaches
concert	One model leads, others refine	Creative tasks, brainstorming, content generation

D-Score Threshold

The dThreshold determines when consensus is considered reached. When the D-score drops below this value, the algorithm stops iterating and returns the result.

0.05 (default) — strong consensus required
0.20 — high confidence, faster results
0.40 — moderate agreement acceptable

Timeout

The client uses a default timeout of 120 seconds (120,000ms) per consensus request. This is configurable via the timeoutMs parameter in the client configuration. The timeout uses AbortController for clean cancellation.