Multi-model deliberation panel. Powered by NaN Builders.
One model has biases. A panel catches blind spots, surfaces contradictions, and builds consensus through diversity. FusionClaw runs multiple AI models in parallel on the same prompt, then a judge synthesizes their answers into one structured analysis.
Deliberation panels are token-heavy. Unlimited tokens make them practical. That's why FusionClaw runs exclusively on NaN Builders — 500M tokens/month per model, no surprise bills, no rate-limit anxiety.
| Problem | Typical API | NaN Builders |
|---|---|---|
| 5-model panel = 5× tokens per prompt | Burns through your quota in hours | 500M tokens/month per model — panels all day |
| Budget anxiety | You hesitate before launching a panel | Launch without budget anxiety |
| Rate limits mid-deliberation | 429 errors break your panel | Generous limits, no throttling |
| Vendor lock-in | Proprietary SDKs | OpenAI-compatible API, drop-in replacement |
| EU data residency | US-hosted, GDPR risk | EU infrastructure |
FusionClaw is the proof. A deliberation panel that spawns 3-5 models per question, each generating 500-2000 tokens, plus a judge that reads all responses. That's 10K-20K tokens per deliberation. Run 100 deliberations a day = 2M tokens/day. That adds up fast on pay-per-token APIs.
With NaN, token cost stops being the limiting factor.
User prompt → Conductor
├── Panel (parallel, different NaN models)
│ ├── Panelist 1: DeepSeek V4 Flash (reasoning)
│ ├── Panelist 2: Qwen 3.6 (balanced analysis)
│ ├── Panelist 3: Gemma 4 (diverse perspective)
│ └── Panelist 4: Mimo V2.5 (lightweight contrast)
│
└── Judge (DeepSeek V4 Flash)
→ Structured JSON analysis
→ Final synthesized answer
- Conductor decides if the task merits deliberation
- Panelists answer the same prompt in parallel, each with a different NaN model
- Judge receives all responses and produces structured analysis: consensus, contradictions, gaps, unique insights, confidence levels
- Conductor synthesizes the final answer from the judge's analysis
| Preset | Panel | Judge | Best for |
|---|---|---|---|
| quality | DeepSeek V4 Flash, Qwen 3.6, Gemma 4 | DeepSeek V4 Flash | Architecture, research, complex analysis |
| fast | DeepSeek V4 Flash, Qwen 3.6 | DeepSeek V4 Flash | Code review, debugging, quick second opinion |
| broad | DeepSeek V4 Flash, Qwen 3.6, Gemma 4, Mimo V2.5 | DeepSeek V4 Flash | Maximum diversity, critical decisions |
| lean | DeepSeek V4 Flash, Qwen 3.6 | DeepSeek V4 Flash | Quick sanity check |
| Model | NaN Endpoint | Strength |
|---|---|---|
| DeepSeek V4 Flash | nan/deepseek-v4-flash |
Fast reasoning, technical depth |
| Qwen 3.6 | nan/qwen3.6 |
Balanced analysis, versatile |
| Gemma 4 | nan/gemma4 |
Google-quality, diverse perspective |
| Mimo V2.5 | nan/mimo-v2.5 |
Lightweight, contrasting angle |
All models available via NaN Builders API — OpenAI-compatible, EU-hosted, 500M tokens/month per model.
Note: The model names above (e.g.
deepseek-v4-flash,qwen3.6) are NaN Builders API identifiers. NaN Builders hosts these models on their EU infrastructure with an OpenAI-compatible API. See nan.builders for the full model list and pricing.
The judge returns structured JSON:
{
"consensus": ["Points all/most models agree on"],
"contradictions": [{"topic": "...", "positions": [...]}],
"coverage_gaps": ["Important aspects no model addressed"],
"unique_insights": [{"model": "...", "insight": "..."}],
"confidence_assessment": {
"high": ["Strong multi-model agreement"],
"medium": ["Partial agreement"],
"low": ["Little agreement or weak evidence"]
},
"recommendation": "Synthesized best answer with uncertainty flags"
}- NaN Builders account and API key
- OpenClaw installed (FusionClaw is an OpenClaw skill)
# 1. Get your NaN API key at https://nan.builders
# 2. Configure in OpenClaw
openclaw config set providers.nan.apiKey "sk-your-nan-key"
# 3. Install the skill
cp -r fusionclaw ~/.openclaw/workspace/skills/import openai
client = openai.OpenAI(
base_url="https://api.nan.builders/v1",
api_key="sk-your-nan-key"
)
# Run 3 models in parallel on the same prompt
import concurrent.futures
prompt = "Should we use microservices or a monolith for a team of 5?"
models = ["deepseek-v4-flash", "qwen3.6", "gemma4"]
with concurrent.futures.ThreadPoolExecutor(max_workers=3) as executor:
futures = {
executor.submit(
client.chat.completions.create,
model=m,
messages=[{"role": "user", "content": prompt}]
): m for m in models
}
results = {f.result().choices[0].message.content: m for f, m in futures.items()}
# Now judge with a final call...That's the whole pattern. FusionClaw just wraps this with structured judging, presets, and safety rails.
Trigger deliberation by:
- Saying "fusion this" or "deliberate on X"
- Asking for a panel analysis
- When your agent detects a task merits deliberation
- Simple questions: If one model can answer it confidently, don't waste a panel.
- Time-critical tasks: Parallel deliberation adds 15-50s of latency.
- Straightforward edits: No need for diverse perspectives on a one-line fix.
- Budget-constrained projects: If you're on a pay-per-token API, the math changes. FusionClaw assumes NaN's flat-rate model.
You can absolutely run a deliberation panel on any provider. But:
- OpenRouter: Pay per token. A 5-model panel with judge = 6 API calls × 1000 tokens average = 6K tokens per question. At GPT-4 prices, that's ~$0.18/question. 100 questions/day = $18/day = $540/month. With NaN: €70/month flat.
- Direct APIs: Same math, more integration work.
- Local models: Free but you need GPUs, and model diversity is limited by VRAM.
NaN Builders hits the sweet spot: enough models for genuine diversity, unlimited tokens for real-world usage, EU-hosted for compliance, OpenAI-compatible for easy integration.
| Setup | Monthly cost | Deliberations/day possible |
|---|---|---|
| OpenRouter (frontier models) | $500+ | ~100 |
| OpenAI direct (GPT-4 class) | $400+ | ~80 |
| NaN Builders (4 models) | €70 | Unlimited* |
| Local (4 models, 2× A100) | $800+ (hardware) | Unlimited |
*Within 500M tokens/month per model fair use. That's roughly 250K deliberations/month under fair use.
MIT — Use it, fork it, sell it. Just point back to NaN Builders.
- NaN Builders — Get your API key
- OpenClaw — Agent framework that runs FusionClaw
- OpenRouter Fusion — The original inspiration