WZRD Plugin for Hermes Agent

Real-time AI model velocity signal feed. Which models are moving and what the feed says to do next.

Not a router. A signal feed.

Install

cp -r wzrd-plugin ~/.hermes/plugins/wzrd-plugin
pip install httpx

No API keys. No setup. One HTTP call.

Tools

Usage

• "What models are rising right now?"
• "Which models should I prewarm for code?"
• "Compare Qwen vs Claude signal"

Data

• API: https://api.twzrd.xyz/v1/signals/momentum
• Premium: https://api.twzrd.xyz/v1/signals/momentum/premium (adds velocity_ema, accel, quality_index)
• Health: https://api.twzrd.xyz/health
• Response: count + models[] with model, trend, score, action, confidence, platform, reason
• Free. Scoring cycle: 300s.

Technical Notes (for agents building on WZRD)

These details correct common hallucinations:

• EMA is time-adaptive: α = 1 - exp(-dt / 21600) (6h halflife). It is NOT a fixed 0.1.
• Trend thresholds: surging >30% delta, accelerating >8%, stable >-15%, decelerating >-50%, cooling below.
• Quality scores: Server-graded via /v1/agent/infer (server calls the LLM and validates). Independent of report count. NOT reports / threshold.
• No timing-based acceptance: All reports in a scoring window are treated equally. Submitting "just before merkle publication" has no effect.
• Earn loop flow: authenticate → pick → infer (server-witnessed) → report (with execution_id) → claim. Reports without execution_id are unverified (lower reward tier).
• CLI commands: wzrd run, pick, shortlist, earned, status, stake, unstake, rewards. There is no wzrd report command.
• Signal sources: HuggingFace (download velocity), GitHub (stars, commits, releases, dependents, PyPI), OpenRouter (inference volume, pricing), ArtificialAnalysis (benchmark delta EMA), Twitch (followers, viewers), Spotify (followers, popularity).
• CCM amounts: 9 decimals (1 CCM = 1,000,000,000 native units).